最終更新: 2026-07-05 (IMU Dead Reckoning 追加 — 102本目、compact baseline)
この文書は、次の AI アシスタントが repo の現在地、最近の差分、次にやるべきことを短時間で掴むための handoff。
最初に本ファイルを読み、その次に:
README.mddocs/paper_ready_reproducibility.mddocs/benchmarks/scan2d/README.md— 2D scan odometry リーダーボード (停止中/背景)docs/index.htmldocs/methods.jsonevaluation/scripts/SETUP_2D_SCAN_BENCHMARK.mdevaluation/src/scan_dogfooding.cppevaluation/scripts/demo_localization_zoo.shevaluation/scripts/generate_demo_report.pyevaluation/scripts/validate_demo_artifacts.pyevaluation/scripts/validate_showcase.pyexperiments/results/index.jsondocs/status_taxonomy.mddocs/budget_profiles.mdevaluation/src/pcd_dogfooding.cppevaluation/src/multimodal_dogfooding.cpp
これが最新・最優先の handoff。
imu_dead_reckoning(102本目) を追加した。 ただしこれは論文再現ではなく compact baseline — LIO / OdoNet / NHC-Net / NN-ZUPT ファミリーの「無補助 IMU dead reckoning 下限」参照実装で、claim tier はdocs/paper_ready_reproducibility.mdの T3 smoke / concept ("This is a compact baseline or concept port.")。新規論文の from-paper 実装ではないため、README の「73 paper reimplementations」は変えていない (breadth の「102 methods」のみ +1)。 中身: static-window init (先頭 2.0 s から gyro bias 推定 + gravity alignment) → midpoint quaternion 姿勢積分 (--imu-dr-eulerで forward Euler に切替可能) → specific force の二重積分。opt-in ablation knob として--imu-dr-zupt/--imu-dr-euler/--imu-dr-no-gyro-bias/--imu-dr-static-init-secを追加。evaluation/src/pcd_dogfooding.cppに methodimu_dead_reckoningとして配線済み (imu.csvが無い dataset では自動 skip)。papers/imu_dead_reckoning/{include,src,test}新設、gtest 5件 (ctest -R imu_dead_reckoningで green):StaticWithKnownBiasHasNoDrift,ConstantYawRateNoTranslation,ConstantWorldAccelerationMatchesAnalytic,ZuptResetsVelocityAndLimitsTailDrift,IntegrateTrajectoryFrameSampling。評価は NCLT 2013-01-10 の 120-frame window のみ (
dogfooding_results/nclt_2013_01_10_120, ms25 IMU ~47 Hz, 軌跡長 11.5 m, ~24 s)。 repro:./build/evaluation/pcd_dogfooding dogfooding_results/nclt_2013_01_10_120 experiments/reference_data/nclt_2013_01_10_120_gt.csv --methods imu_dead_reckoning。 結果 (ATE m / RPE %/100m): default (pure DR) 9.071 / 170.617 (zupt_frames=0)、--imu-dr-zupt2.887 / 30.607 (zupt_frames=481/1051 — この短い window は停止区間の 割合が大きいため ZUPT の ATE 改善は継続走行を代表しない、over-claim 禁止)、--imu-dr-euler10.280 / 198.899、--imu-dr-no-gyro-bias24.676 / 482.028 (全 ablation 中で最大の劣化 — static-init gyro-bias 推定が最も load-bearing)。 同じ window での family context: OdoNet 138.872 / 1842.230 (KITTI Raw OXTS 学習済み CNN 重みが NCLT の ms25 IMU に対して out-of-domain)、 NHC-Net 4.295 / 102.938、NN-ZUPT 3.799 / 96.164。 matrix manifest:experiments/imu_dead_reckoning_nclt_2013_01_10_matrix.json、 aggregate:experiments/results/imu_dead_reckoning_nclt_2013_01_10_matrix.json。run_experiment_matrix.pyによりdocs/experiments.md/docs/decisions.md/docs/interfaces.md/experiments/results/index.jsonは additive に更新済み。 注意: runner の自動 heuristic がdocs/decisions.md上でzuptvariant を "Adopt as current default" とラベルしたが、これは shared ATE/RPE スコアだけを見た 機械的判定であり、手法自体の実際の default は pure DR (ZUPT off) のまま変わらない (module README にその旨を明記)。 full NCLT session の書き出しは、評価時点でディスク残 18 GB / 使用率 97% に対し 約 3.6 GB 必要になるため見送った (120-frame window のみ)。2026-07-05 追記: full-session (5105 frames, ~17分/1021.7s, 1138.8 m) 評価を実施した。 ローカル disk が依然 18 GB しか無いため、
/media/sasaki/aiueo(外付け SSD, 1.6 TB free) にdogfooding_results/nclt_2013_01_10_full(実測 6.3 GB、想定より大きい) を書き出し。experiments/reference_data/nclt_2013_01_10_full_gt.csvは既存ファイルと byte-identical (md5ff32d5666754fc1fb95333a3835752f4) だったため差分なし。IMU-DR は点群を読まない (imu.csv + frame_timestamps.csv のみ使用) ため、full-session 4 variant (family run 込み) の実行は数秒〜2分で完了 (export 自体は 6m45s)。結果 (ATE m / RPE %/100m): default 288700.449 / 67435.005 (zupt_frames=0、軌跡長の約250倍)、--imu-dr-zupt14531.743 / 2859.304 (-94.97%/-95.76%、zupt_frames=3984/48122≈8.3% — 120-frame window の ~46% より大幅に低いが、それでも ZUPT が drift の大半を除去。 「stationary 時間が長いから効く」ではなく「稀なリセットで速度誤差の蓄積を打ち切る」効果と判明)、--imu-dr-euler291892.627 / 68484.744 (+1.11%/+1.56%、依然として二次的な効果)、--imu-dr-no-gyro-bias672302.751 / 139392.868 (+132.9%/+106.7%、全 ablation 中で 最大の劣化、120-frame window と同じ順位)。4 run とも NaN/オーバーフロー無しで有限値 (hundreds-of-km は非現実的な bug ではなく、無補正 gyro heading drift が ~17分で複利的に 蓄積した honest な結果)。manifest:experiments/imu_dead_reckoning_nclt_2013_01_10_full_matrix.json、 aggregate:experiments/results/imu_dead_reckoning_nclt_2013_01_10_full_matrix.json、run_experiment_matrix.py --merge-existing-indexで additive に反映 (generate_leaderboard.py は未使用)。 注意: full-session の pcd_dir は removable media (/media/sasaki/aiueo) 上にあり、 ドライブが未マウントだと再現できない — 再エクスポート手順は module README (papers/imu_dead_reckoning/README.md) の Limitations に記載。ドキュメント更新:
papers/imu_dead_reckoning/README.md(新規)、docs/methods.jsonにIMU Dead Reckoningエントリ追加 (scope: Compact baseline, signals: [IMU])、 README.md の badge/breadth カウントを 101→102 に更新 (paper reimplementations の 73 は不変)。experiments/families.jsonはあえて変更していない — 同ファイルには そもそもnn_zupt/odonet/nhc_net(IMU dead reckoning ファミリー) のエントリが 存在せず、budget-profile 付き KITTI/LIO/VIO ファミリーだけを対象にした registry (docs/budget_profiles.md/docs/status_taxonomy.mdから参照) と見られるため、 前例のない family を追加すると逆に registry の一貫性を崩す。次に触るときは このファイルの scope をまず明確化すること。次にやるとよいこと: (1) [2026-07-05 完了] NCLT full-session (外付け SSD 経由) で 継続走行での ZUPT 効果を再検証済み — 上記追記参照、 (2) [2026-07-05 完了] 第二データセットとして KITTI Raw OXTS の
imu.csvをevaluation/scripts/kitti_oxts_imu_for_dogfooding.pyで用意し評価済み — 下記追記参照、 (3) OdoNet/NHC-Net/NN-ZUPT との DR family 比較を README かdocs/paper_ready_reproducibility.md側の比較表に昇格するか検討する (現状は module README 止まり)、(4) full-session の family run (OdoNet/NHC-Net/NN-ZUPT) も同時に測定済み (OdoNet ATE 1397.891 m / RPE 753.600%, NHC-Net 279.794 m / 84.226%, NN-ZUPT 257.397 m / 82.813%; family JSON は SSD 上のfull_family.jsonに保存のみで repo にはコミットしていない — 必要なら別 matrix manifest化を検討)。2026-07-05 追記: KITTI Raw drive 2011_09_26_0009 (OXTS) を第二データセットとして 評価した。
s3.eu-central-1.amazonaws.com/avg-kittiの公開・無認証ミラーから sync (1.79 GB) + calib (4 KB) を/media/sasaki/aiueo/loc_zoo/kitti_raw_downloadにダウンロードし、kitti_raw_to_benchmark.py --write-imu-csvで 200-frame window と full (443 frames) を/media/sasaki/aiueo/loc_zoo/dogfooding_results/kitti_raw_0009_{200,full}にエクスポート。GT は既存コミット済みexperiments/reference_data/kitti_raw_0009_{200,full}_gt.csvと md5 一致 (差分なし)。OXTS レート: sync パッケージは 1 Velodyne フレームにつき 1 OXTS パケット (447 サンプル/443 フレーム、~46.2s) = 実効 ~9.7 Hz のみ (100Hz は 別の_extractパッケージのみ)。100Hz variant は検討したが、frame_timestamps.csvのtimestamp列が Velodyne ファイルの位置インデックスであり、GT 対応もそのインデックス を直接使うため、imu.csvを実秒に変換すると GT 対応が壊れる — 既存のkitti_raw_0009*系の全リーダーボードエントリに影響するため見送り、正直に理由を記録。 さらに深刻な単位ミスマッチ:imu.csvのスタンプも同じ位置インデックス上にあり (1.0 単位間隔)、harness はFrame gap [s]: min=median=mean=max=1.000と報告する — strapdown 積分の dt が (max_dt=0.5s でクランプされた) 0.5s/step になり、実際の ~0.103s/step の約4.8倍。static-init window (2.0s 設定) も実質 ~3 サンプル (~0.3s 実時間) にしかならない。static-init 前提の破綻: drive 0009 は t=0 で 既に ~10.7 m/s で巡航中 (200-frame window では最低速度 1.339 m/s、一度も静止せず)。 gyro std ゲート (0.05 rad/s) はこの区間の低ヨーレートのため発火せず、並進運動による 静止仮定違反を検知できない盲点を確認。ZUPT 誤検出: 200-frame window で ZUPT が 174/199 (~87%) 発火するが実速度は一度も 0.5 m/s を下回らず (ほぼ全て偽陽性)、 full でも 344/442 (~78%) 発火 vs 実静止割合 ~9.7% (43/443)。それでも ZUPT は両 window で最大の改善効果 (200f: -97.81%/-97.72%, full: -93.51%/-91.28%)。 gyro-bias reversal (新知見): NCLT では no-gyro-bias が最大の劣化要因だったが、 KITTI Raw 0009 では逆に改善 (200f: -87.36%/-86.49%, full: -87.15%/-87.46%、両window で再現) — ~3サンプルの static-init window が巡航中の実角速度で汚染されるため。 family context (OdoNet/NHC-Net/NN-ZUPT) は in-domain (KITTI-OXTS 訓練済み) で IMU-DR を 1-2桁上回る (200f: 856/122/123m, full: 1723/181/186m)。ただし訓練に drive 0009 自体が含まれていたか記録が無く (val drive は 0056 のみ既知)、 train/eval overlap の可能性は否定できないため正直に記録。副次発見: drive 0009 の sync パッケージは Velodyne フレーム 177-180 (native index) が欠落しているが OXTS は欠落なし (447 vs 443) —kitti_raw_to_benchmark.pyが両方を native frame number ではなく position で index するため、position 177 以降 (~60%) は GT pose と point cloud が ~4 フレームずれるという既存 (今回発見、今回導入ではない) バグを 発見。IMU-only 手法 (imu_dead_reckoning/odonet/nhc_net/nn_zupt) には影響しないが (GT と imu.csv は共に OXTS-native で自己整合)、kitti_raw_0009*を使う全 scan-matching 手法のリーダーボードに影響しうる — 共有 exporter の修正は本 pass の scope 外なので 次に触る人向けにpapers/imu_dead_reckoning/README.mdの Limitations に記載のみ。 manifest:experiments/imu_dead_reckoning_kitti_raw_0009_matrix.json(200-frame),experiments/imu_dead_reckoning_kitti_raw_0009_full_matrix.json(full)、run_experiment_matrix.py --merge-existing-indexで additive に反映 (generate_leaderboard.py は未使用)。LiTAMIN2 NTNU LiDAR degeneracy health (2026-06-13 時点) 以降は依然有効な背景 (paper-ready reproducibility hardening 方針、continuous-time LIO tuning 履歴) として 以下に残す。
2026-06-13 現時点のアクティブ方向は paper-ready reproducibility hardening。 ユーザ指示「2d ha ittan iran!」により、2D LiDAR scan odometry は一旦停止。 新規手法追加より、既存の 101 手法を claim tier で分け、論文で主張できる T0/T1 subset と adapter / compact baseline を明確に分離する。README は breadth を見せるが、manuscript-level claim は
docs/paper_ready_reproducibility.mdに従う。 直近は LiTAMIN2 の論文再現ギャップを詰めるため、papers/litamin2/src/litamin2_registration.cppとevaluation/src/pcd_dogfooding.cppに covariance-gradient / covariance floor / line-search の opt-in knob を追加し、KITTI Odometry seq02/05/07/08 full の correspondence/covariance matrix をexperiments/pending/とexperiments/results/に追加した。 raw--litamin2-covariance-gradientは seq02 full で ATE 384.474 m / RPE 5.679 % まで崩れ、 幾何平均 RPE も 1.451 % に悪化。--litamin2-covariance-gradient-weight 0.1と--litamin2-line-searchは崩壊を止め、seq02/05/07/08 幾何平均 RPE を baseline 0.80645 % から 0.80581 % へ微改善したが、seq02 ATE が 50.622 m から 81.961 m に悪化するため paper default には未昇格。続けて--litamin2-coarse-to-fine-voxels 3.0,2.0,1.0も 実装・full 評価した。108-frame seq02 smoke は RPE 11.969 % から 0.670 % へ大きく改善したが、 full seq02/08 が悪化し、seq02/05/07/08 幾何平均 RPE は 1.185 % に悪化した。 追加で LiTAMIN2 のrefresh_interval/map_radius/map_max_points/ seed acceptance gate を CLI sweep 可能にした。短窓では refresh1 が seq07 に効くが、 full seq02 では--litamin2-refresh-interval 1が RPE 4.412 %、tight seed gate が RPE 2.306 % で、coarse-to-fine default の 2.348 % から十分には戻らない。 さらに--litamin2-coarse-to-fine-iterationsで段ごとの反復数を sweep 可能にした。 full seq02 は3,3,6で ATE 48.589 m / RPE 1.224 % / 89.5 FPS、2,2,8で ATE 55.516 m / RPE 0.976 % / 88.0 FPS まで戻るため、粗段の over-iteration が long drift の主因候補。ただし2,2,8を full seq05/07/08 に 横展開すると ATE は大きく改善する一方で RPE は seq05 0.630 %、 seq07 0.782 %、seq08 1.386 % に悪化し、seq02/05/07/08 幾何平均 RPE も baseline 0.806 % から 0.903 % に悪化した。ATE/RPE trade-off が強いため default にはしない。upstream 非公式実装も確認し、KITTI例は直前scanをtargetにする scan-to-scan + relative pose accumulation で、dogfooding の scan-to-map + GT/velocity seed + local map refresh とは運用が違う。covariance shape cost の更新項は upstream point-to-voxel でもコメントアウトされており、現defaultのoptimize_covariance_cost=falseは整合。--litamin2-scan-to-scanも opt-in 追加済み。 seq02 108-frame は ATE 1.395 m / RPE 2.568 % (GT relative seed)、 ATE 1.474 m / RPE 2.774 % (no-GT) で動くが、full seq02 は ATE 445 m / RPE 21.8-24.4 % までdriftするため default にはしない。 軌跡一貫性を見る--litamin2-max-motion-translation-delta/--litamin2-max-motion-rotation-deltaも opt-in 追加済み。coarse-to-fine default に0.25 / 0.05を足すと seq02/05/07/08 full ATE は 0.957 / 0.957 / 0.714 / 0.978 m まで下がるが、RPE は 1.352 / 1.321 / 1.111 / 1.359 % で、幾何平均 RPE は baseline 0.806 % から 1.282 % に悪化。ATE重視診断には有用だが RPE default にはしない。 今回さらに scan-to-map local-map policy を CLI sweep 可能にした。--litamin2-local-map-policy refresh|accumulate|keyframeを追加し、defaultrefreshは 従来互換、accumulateは全 frame を rolling map に貯め target 再構築だけを間引く、keyframeは--litamin2-keyframe-translation/--litamin2-keyframe-rotation到達時に map 追加と target 再構築を行い、periodic refresh でも同じ処理を行う。 KITTI seq02 108-frame の coarse-to-fine smoke は refresh ATE 0.556 m / RPE 0.683 %、accumulate 0.554 / 0.698、keyframe 0.526 / 0.692。 full seq02 は2,2,8 + keyframeが ATE 1.245 m / RPE 1.208 % となり、2,2,8 + refreshの ATE 55.516 m / RPE 0.976 % に対して ATE/RPE trade-off がさらに明確。 seq02/05/07/08 full の2,2,8 + keyframeは ATE 1.245 / 1.522 / 0.812 / 1.317 m、 RPE 1.208 / 0.762 / 0.635 / 1.456 %、幾何平均 RPE 0.961 %。 ATE診断用の有力 knob だが、baseline RPE 0.806 % と2,2,8 + refreshRPE 0.903 % に 届かないため default にはしない。 ユーザ提案のhttps://github.com/ntnu-arl/lidar_degeneracy_datasets評価にも着手。 既存のevaluation/scripts/SETUP_LIDAR_DEGENERACY_BENCHMARK.md/experiments/results/lidar_degeneracy/pipeline に LiTAMIN2 を追加するため、evaluation/src/litamin2_window_odometry.cppを新設し、evaluation/scripts/run_lidar_degradation_health.py --method litamin2とevaluation/scripts/summarize_lidar_degeneracy_health.pyに配線した。 既存抽出済みの NTNU selected windows で GT-free health check を実行済み:fog_200は 3/3 selected windows accepted/converged、policy pass 3/3、 max used path 1.036 m。tunnel_geom_2700_200は 4/4 accepted/converged、 policy pass 4/4、max used path 3.035 m。 追加で stress sensitivity sweep を実行し、experiments/results/lidar_degeneracy/litamin2_stress_sensitivity/summary.mdに集計した。variant は default /--litamin2-icp-only/--litamin2-correspondence-search-radius 1/ seed gate 0.5 m・0.1 rad / step gate 0.5 m・10 deg。ICP-only、tight seed、tight step は default と同じく fog/tunnel 全 selected windows accepted/converged で health flag なし。radius1 は tunnel の max path を 3.035 m -> 2.810 m に下げたが、fog nominal baseline にlow_used_pathfalse alarm を1件作るだけで、stress-only failure signal にはならなかった。 これは GT無しの短窓 health であり paper-level accuracy claim ではない。 次は LiTAMIN2 の degeneracy health を README / paper-ready docs に「GT-free robustness probe」として 位置づけるか、外部 GT / pose source が取れる dataset で実誤差 calibration へ進むのが妥当。§0 (2026-06-02 の OSS Showcase) 以降は依然有効な背景 (showcase/demo/CI、3D benchmark 履歴、 recipe 由来) で、2D の詳細は §00.6c〜§00.66 を背景として読むこと。
- 方針転換: ユーザ指示「しばらくは 2D LiDAR でやっていく」→
scan_dogfoodingharness 新設。- papers 43–50 完了 (RF2O / PL-ICP / CSM / Kinematic-ICP / PSM / NDT-2D / IDC / MbICP)。
- 公開 dataset: Bonn 2D-SLAM JSON →
intel_val_73,fr079_val_384,mit_val_33を repo に commit。- 合成 fixture:
rf2o_smoke(60f),rf2o_corridor(120f slow motion)。- CSM 改善 (commit
3fc5be0): distance transform + 3-level pyramid → fr079 drift 38.9%→20.6%。- MbICP (50本目): config-space metric ICP を追加し、canonical JSON を8法で refresh。
- ドキュメント整備:
docs/benchmarks/scan2d/README.mdハブ、 README / SETUP /public_bundle.json/ 各 paper README を 8 法に同期。- git (当時):
mainはorigin/mainより 2 commit ahead (361a592IDC,3fc5be0CSM-DT)。 markdown 整理分はワークツリーに未 commit だった。
- 50-star 施策 PR #11/#12 マージ: hero GIF、誤差ヒート版 seq07 ギャラリー、social card、README -28%。
- 32 本目 I-LOAM (PR #13): 強度 ablation positive; mapping 統合で seq00 0.899% drift。
- 33–40 本目: PL-LOAM / InTEn-LOAM / MCGICP-LO / ICPSC-LO / VLOM / OdoNet / NHC-Net / NN-ZUPT。
- 41–42 本目: FR-LIO (NCLT 0.63% drift) / PG-LIO (NCLT honest negative → 保留)。
- Intensity / LiDAR-visual / IMU 3 カテゴリ shortlist 完了 (多く honest negative、正直記録)。
ユーザの一貫した指示:
- 各 "tugi" / "tudukete" / "dondon ikou" で 次の 1 論文 を実装する (自律的に新規論文を 始めない — 明示トリガが必要)。候補が尽きたら新規 web サーベイを行う。
- 対象は 著者の公開実装が無い 論文のみ (OSS 済みは除外)。KITTI で評価可能 な 古典・幾何・確率的手法 (deep/neural はスコープ外、CV 予測 front-end に落ちる)。
- 結果は 正直に 反映する (near-redundant / honest negative も隠さず記録)。
- 機構ユニットテスト 3 件 + KITTI full seq00/seq07 評価 + README leaderboard 行 +
docs/methods.jsonエントリ + module README + memory 追記 + clean commit が 1 論文の単位。 - コミットは自分名義のみ (Co-Authored-By を付けない)。PR/コミットに AI 生成表記を 入れない (CLAUDE.md 制約)。応答は日本語。
from-paper キャンペーンと並行して、ユーザ指示「50 star wo mezasu」の発見性・第一印象・
信頼シグナルを上げる視覚/ドキュメント資産を整備し main にマージ済み。コミットは
自分名義のみ (AI 表記なし)。
- README hero アニメ GIF
docs/assets/hero_seq00.gif(+.mp4): KITTI seq00 上面視で KC-LO/KISS-ICP/TrICP-LO を GT 上に時間方向に描画。生成器evaluation/scripts/animate_trajectory.py(入れ子バンド描画で全色が縁取りで見える)。 - seq07 軌跡ギャラリー
docs/assets/grid_seq07.png: 当初は 15 手法の生軌跡 small-multiples だったが、優秀手法が全部 GT に重なって差が見えないため、PR #12 で各軌跡を GT からの フレーム毎の距離で色付けする誤差ヒート版に刷新 (turbo, 3m clip, cool=追従/warm=ドリフト)。 生成器evaluation/scripts/plot_trajectory_diff.py(--mode heat/--mode curve)。 旧生成器plot_trajectory_grid.pyも残置。 - social card + OGP
docs/assets/social_card.png(1280×640): 生成器evaluation/scripts/make_social_card.py。docs/index.htmlの og:image/twitter:image に配線。 - community health files:
CONTRIBUTING.md/CITATION.cff/.github/ISSUE_TEMPLATE/*/.github/PULL_REQUEST_TEMPLATE.md。 - README -28% 削減 (561→428 行)、Contributing/Citation 節追加。
- CI ゲート整合: hero/social card 差し替えで
evaluation/scripts/validate_showcase.pyの スニペット契約が壊れ CI が落ちた → 新資産 (hero_seq00.gif の GIF 署名チェック等) に合わせて 修正済み。README/index.html を触ったらpython3 evaluation/scripts/validate_showcase.py --root .を必ず回すこと (CI と同じ契約)。 - GitHub About 変更 (
gh repo edit --description): 「with ROS 2」を外し from-paper を前面に → "C++ from-paper reimplementations of LiDAR localization & odometry papers — honestly benchmarked on KITTI, with tests." (topics のros2タグは検索性維持のため残置)。
PENDING — 手動・ユーザのみ (コード不可): GitHub repo Settings → General → Social
preview に docs/assets/social_card.png をアップロード。未設定だとリンク共有時に汎用カード。
詳細 memory: loc-zoo-50star-campaign。
| Item | Value |
|---|---|
| Branch | main |
vs origin/main |
ahead 3 after KC-LO ablation commit if not pushed |
| 実装済み from-paper 論文数 | 60 本 (3D 再開: Mesh-LOAM + ELO + ID-LIO + RF-LIO + TC-LVGF + OPL-LVIO + V-LOAM2015 + TC-VLO + AD-VLO + TC-MVLO; 2D papers 43–50 は停止中) |
docs/methods.json |
101 手法 |
| 2D scan matchers | 8 法 — rf2o,pl_icp,csm,kinematic_icp,psm,ndt_2d,idc,mb_icp |
| 2D fixtures (committed) | 5 — intel/fr079/mit (Bonn) + rf2o_smoke + rf2o_corridor |
| 2D リーダーボード hub | docs/benchmarks/scan2d/README.md |
| 直近完了 | Quadric-LO plane-fallback ablation — seq00/07 fallback on/off raw JSON + bundle 4-method化 |
| 直近完了 (3D) | RF-LIO (101手法目) — removal-first dynamic LIO、KITTI seq00/07 full 完走 |
| その前 (3D) | V-LOAM2015 / TC-VLO / AD-VLO / TC-MVLO (97-100手法目) — LiDAR-visual adapter family、KITTI seq00/07 full 完走 |
| 2D 直近 (停止中) | MbICP (50本目) + 8-method canonical benchmark refresh |
| PG-LIO (42本目) | NCLT honest negative → 保留 (§00.52) |
直近コミット列 (main、2D 部分):
3fc5be0 Improve CSM with distance transform and multi-resolution pyramid search. ← local only
361a592 Add paper 49 IDC dual-correspondence 2D scan odometry. ← local only
8ea40f1 Add paper 48 NDT-2D scan odometry with public 2D benchmark results. ← origin/main
a2817ff Add fr079 and MIT public 2D scan fixtures with multi-dataset benchmarks.
709d25c Add paper 47 PSM and Intel Lab 2D scan benchmark with CI smoke.
0a9f0ad Add 2D scan eval infra: drift metrics, bag prep, and corridor fixture.
8339329 Add paper 46: Kinematic-ICP 2D unicycle scan odometry with wheel prior.
5e2147c Add paper 45: CSM correlative 2D scan matching with scan_dogfooding.
1c94e71 Add 2D scan odometry (RF2O, PL-ICP), scan_dogfooding, and PG-LIO baseline.
3D campaign 部分 (参考):
9184524 Add papers 35-36: MCGICP-LO and ICPSC-LO intensity odometry.
23504ff Add paper 41: FR-LIO RC-Vox LIO with sub-frames and ESKS.
8a980fc Add IMU papers 38-40, RGB LiDAR-visual eval, InTEn mapping, cross-dataset benchmarks.
注: §00.53〜§00.56 の番号は 3D 37〜40 本目 (VLOM/OdoNet/NHC/NN-ZUPT) と 2D 43〜46 本目 (RF2O/PL-ICP/CSM/Kinematic-ICP) で重複している。 2D の詳細は §00.6c 以降 を正とする。
整備済み・IMU 無し・CV 予測が効く KITTI では:
- robust / soft 機構は near-redundant: Student-T (重尾)、GMM (soft assignment)、 GNC (継続 TLS)、MCC (相関エントロピー) はいずれも front-end が ~KISS-ICP の point-to-plane に退化。外れ値が少なく重みがほぼ一様になるため。機構の正しさは ユニットテストで担保し、KITTI では「near-redundant」と正直に記録。
- 退化判定・粒子フィルタ・direct/range-image はKITTIで沈黙 or 有害: DiLO (direct) は full seq で発散 (18-19%)、Spectral-LO (ICP-free BEV phase-correlation) は最速 ~14 FPS だが粗い (~12-14%)、UA-LIO/DegenSense は未競争。
- point-to-plane 主体が低 drift: point-to-line / 分布ファクタは drift 増。地面拘束は long-seq の局所 drift (RPE) を抑えるが ATE は悪化しうる (RPE↓/ATE↑ split)。
- 退化しきい値は並進ブロック相対基準が鉄則 (絶対しきい値は移植不可で発散)。
- seq00 RPE リーダーは M-GCLO 0.835% (multiple-ground-plane、ただし ATE 19m の split)。 top9 (M-GCLO〜Adaptive-ICP) は両 seq で KISS-ICP に match/beat。
- 暗黙的曲面 (IMLS) も point-to-plane に退化: IMLS-SLAM(29) のボクセルダウンサンプル 局所マップ上の implicit MLS 曲面は実質 point-to-plane となり KISS-ICP を僅かに下回る (honest negative)。原論文の 0.4–0.7% drift は生 HDL64 密度前提。観測性サンプリング は有効で、~800 点でも全点と同精度を保ち FPS は 2 倍。
- LTS トリミングの自動オーバーラップ推定は KITTI で収縮: TrICP-LO(30) の FRMSD は ノイズで残差が連続分布する KITTI では ξ を下限 (0.8) まで縮め続け、実質「最良 80% 分位 の固定トリム」= ロバスト分位 point-to-plane として動作。それでも competitive で seq00 ATE 10.0 m は KISS-ICP 12.0 m を上回る (順位トリミング機構自体はユニットテストで有効)。
- 対応点を取らない密度相関 (KC) は KITTI で恒例パターンを破り positive: KC-LO(31) の カーネル相関 (Renyi 二次エントロピー) + σ アニーリングは両 seq で KISS-ICP の drift を 下回り、fixed-sigma ablation 後の seq07 RPE 0.510% は leaderboard 全体トップ。 離散最近傍対応のノイズに頑健な soft point-to-point が効く。KITTI + CV 予測では fixed σ=0.4 が annealed σ 1.5→0.4 と同等精度で 1.9-2.2x 速い。
- 反射強度は KITTI でも対応の曖昧性解消に効く (intensity ablation positive): I-LOAM(32)
の強度拡張対応 + 強度重みは、同一の幾何パイプライン (強度 OFF) 比で drift を両 seq -18〜20%、
ATE を最大 -35% 改善。KITTI の未校正・粗い 8-bit 強度でも、強度を主信号ではなく対応コスト
項 + 残差重みとして使えば壊れず効く。mapping 統合後 (aloam LaserMapping, デフォルト ON)
seq00 0.899% / seq07 0.575% → from-paper leaderboard 同列 (A-LOAM ~0.61% 水準)。
強度 ablation は
--i-loam-no-mappingで scan-to-scan のみに分離して維持。
各論文の機構・結果・教訓は per-paper memory と README / docs/benchmarks/scan2d/README.md を参照。
2D scan odometry (papers 43–50) の追加知見:
- fixture 依存が支配的 — RF2O@Intel, PSM@fr079, PL-ICP@corridor。単一手法の「勝者」は存在しない。
- scan-to-scan 限界 — IDC/PSM/PL-ICP/MbICP/NDT-2D/Karto は robot-frame local map 済。RF2O (P18) と Kinematic-ICP (P19) は opt-in — 前者は projection 型で long 窓全敗、後者は point 型でも trade-off に safe config 無し。これで 2D 全 9 法の local map 調査完了。長 Bonn log では drift 15–30% 帯が典型。
- projection 型 local map の罠 (P18) — 点群を min-range polar profile に再投影する reference (RF2O local map; IDC/PSM も同族) は val 窓で効くが long train 窓で爆発 (fr079_train_200: IDC 85% / PSM 681% / RF2O 99%)。bin 量子化 + 近距離バイアスが原因で、age 減衰でも救えない。
- 合成 corridor の罠 — PL-ICP 0.4% でも fr079 41% — synthetic 成功は real 一般化の証拠にならない。
- CSM engineering win — DT + pyramid で fr079 38.9%→20.6%。corridor ~73% は honest negative 継続。
- correspondence-free 2D — NDT-2D は Intel で RF2O 近傍 (14.8% vs 14.3%) だが corridor 22%。
- metric-space ICP — MbICP は fr079 16.6%、MIT 27.3%、corridor 0.46%。fixture winner ではないが PL-ICP より real logs で安定し、corridor では PL-ICP に近い。
- wheel odom 依存 — Kinematic-ICP は
--wheel-odom-from-gt必須。real encoder 無しでは実用度低。 - GT proxy 限界 — Bonn GT は scan-matched odometry。centimeter truth ではない — drift は相対比較用。
3D Velodyne / multimodal (papers 1–42, 代表ファミリ)
- Ground / multi-plane: M-GCLO(20), DAMM-LOAM(11), DALI-SLAM, Terrain-RBF-LIO, NHC-LIO(23)
- Voxelmap / surface: CT-VoxelMap, R-VoxelMap, Quadric-LO(21), CUBE-LIO, FR-LIO(41)
- Implicit surface / MLS: IMLS-SLAM(29) — honest negative
- Robust estimation: Student-T-LO(24), GNC-LO(27), MCC-LO(28), TrICP-LO(30), SVN-ICP, Adaptive-ICP
- GMM / EM: GMM-LO(26)
- Spectral / phase: Spectral-LO(25)
- Correlation / entropy: KC-LO(31) — positive (seq07 RPE トップ)
- Intensity / reflectivity: I-LOAM(32), InTEn-LOAM(34), MCGICP-LO(35), ICPSC-LO(36), PL-LOAM(33), VLOM(37)
- IMU / learning: OdoNet(38), NHC-Net(39), NN-ZUPT(40), PG-LIO(42, 保留)
- Distribution / factor: PCR-DAT(18), LiDAR-IBA, D2-LIO
- Direct / range-image: DiLO(22) — honest negative; PG-LIO photometric — honest negative
2D planar scan odometry (papers 43–50, scan_dogfooding)
- Range flow (dense, no correspondences): RF2O(43)
- ICP family: PL-ICP(44), Kinematic-ICP(46, wheel prior), MbICP(50, config-space metric)
- Correlative / grid: CSM(45, DT+pyramid), NDT-2D(48, Gaussian cells)
- Polar / dual correspondence: PSM(47), IDC(49, CP+RR fusion)
papers/imls_slam/ — 暗黙的移動最小二乗 (IMLS) 曲面への scan-to-model マッチング。
出典: J.-E. Deschaud, ICRA 2018 (arXiv:1802.08633)。著者は後発の CT-ICP は公開したが
IMLS-SLAM 本体は未公開。機構: 過去スキャンの有向点群が定義する暗黙的曲面
I^P(x)=Σ W_i (x−p_i)·n_i / Σ W_i, W_i=exp(−||x−p_i||²/h²) への符号つき距離 (point-to-
implicit-surface) を Gauss-Newton で最小化。さらに 6-DoF を最も拘束する点を選ぶ観測性ベース
サンプリング (並進 |n·e| / 回転 |(p×n)·e| の 6 リスト上位)。単一最近傍平面の
point-to-plane や voxel-surfel とは別系統の幾何当てはめ。
- 統合済み: header/src/test (3/3 pass)、root+evaluation CMake、
pcd_dogfooding8 スポット (include / method-check / DogfoodingOptions struct / runImlsSlam / options var / arg-parse--imls-slam-{fast,dense}-profile/--imls-slam-h/--imls-slam-no-sampling/ dispatch / method-list 文字列 2 箇所)、docs/methods.json(69手法)、module README、seq00/07 ベンチ JSON。 - 結果: seq00 RPE 1.000% / ATE 17.6 m (7.4 FPS)、seq07 RPE 0.700% / ATE 3.04 m (9.3 FPS) — honest negative: KISS-ICP (同データ 0.872% / 0.618% を完全再現) に両 seq で僅差で劣る。 整備済み KITTI ではボクセルダウンサンプル局所マップ上の implicit MLS が実質 point-to-plane に 退化。原論文 0.4–0.7% drift は生 HDL64 密度前提。観測性サンプリングは有効 (seq07 全点 2023 点 0.701% vs サンプリング 807 点 0.700%、FPS 2 倍)。
- 完了: seq00/07 ベンチ JSON、README leaderboard 行 (GNC-LO と CT-VoxelMap の間)、 module README、methods.json (69手法)、memory 追記、commit。
- ※ この環境は archive copy で KITTI 点群・Ceres を欠く。調達手順は memory
loc-zoo-archive-build-env参照 (Ceres をthird_party/ceres/installへ source build、 KITTI velodyne をdata/kitti_rawへ再 download)。
papers/tricp_lo/ — Trimmed / Fractional ICP (最小トリム二乗, LTS) point-to-plane。
出典: D. Chetverikov et al., ICPR 2002 / IVC 2005 + J. Phillips et al., 3DIM 2007 (FICP)。
対点群レジストレーション論文でオドメトリ公開実装は無い。機構: 残差を順位でソートし
最良 ξ 割合のみを最小二乗に使う高破壊点ロバスト推定。ξ は FRMSD(ξ)=sqrt(e(ξ))/ξ^λ の
最小化で自動推定 (FICP)。M 推定/GNC の magnitude 重み付けとは別系統 (rank-based)。
- 統合済み: header/src/test (3/3 pass)、root+evaluation CMake、
pcd_dogfooding8 スポット (include / method-check / DogfoodingOptions struct / runTricpLo / options var / arg-parse--tricp-lo-{fast,dense}-profile/--tricp-lo-overlap/--tricp-lo-fixed-overlap/--tricp-lo-lambda/ dispatch / method-list 文字列 2 箇所)、docs/methods.json(70手法)、 module README、seq00/07 ベンチ JSON。 - 結果: seq00 RPE 0.931% / ATE 10.04 m (5.6 FPS)、seq07 RPE 0.662% / ATE 2.00 m (6.2 FPS) — competitive。KISS-ICP (同データ 0.872%/0.618%) に RPE で僅差で劣るが seq00 ATE 10.0 m は KISS の 12.0 m を上回り、IMLS(29)・robust 群の多くより良好。
- FRMSD 自動推定の所見 (honest): 推定 ξ は両 seq とも下限 0.800 に張り付く。整備済み KITTI ではノイズで残差が連続分布し FRMSD が良点を非オーバーラップと誤認して ξ を縮め 続けるため、実質「最良 80% 分位の固定トリム」= ロバスト分位 point-to-plane として動作。 LTS の順位トリミング機構自体はユニットテストで有効。
- 完了: seq00/07 ベンチ JSON、README leaderboard 行 (SVN-ICP と GMM-LO の間)、 module README、methods.json (70手法)、memory 追記、commit。
papers/kc_lo/ — Kernel Correlation (対応点を取らない / correspondence-free) 位置合わせ。
出典: Y. Tsin & T. Kanade, ECCV 2004。対点群レジストレーション論文でオドメトリ公開実装は
無い。機構: 変換後スキャンとモデルの総親和度 C(θ)=ΣΣ exp(−||Tx_i−y_j||²/2σ²) を最大化
(= 結合密度の Renyi 二次エントロピー最小化)。各点を近傍モデル点の親和度重み平均へ引き寄せる
soft point-to-point を σ アニーリング (粗→細) で解く。ICP の最近傍対応・GMM-EM の潜在割当・
NDT のボクセル離散化のいずれも用いない新ファミリ。
- 統合済み: header/src/test (3/3 pass)、root+evaluation CMake、
pcd_dogfooding8 スポット (include / method-check / DogfoodingOptions struct / runKcLo / options var / arg-parse--kc-lo-{fast,dense}-profile/--kc-lo-sigma/--kc-lo-sigma-init/--kc-lo-no-anneal/ dispatch / method-list 文字列 2 箇所)、docs/methods.json(71手法)、module README、 seq00/07 ベンチ JSON。 - 結果 (best fixed-sigma profile): seq00 RPE 0.837% / ATE 13.40 m / 2.65 FPS、 seq07 RPE 0.510% / ATE 0.86 m / 3.12 FPS — positive: 両 seq で KISS-ICP (同データ 0.872%/0.618%) の drift を上回り、seq07 RPE 0.510% は leaderboard 全体トップ (旧 Adaptive-ICP 0.569% 超え)。seq00 は RPE↓/ATE↑ split。 恒例の「全機構が point-to-plane に退化」を破る数少ない例。
- sigma schedule ablation: annealed σ 1.5→0.4 は seq00 0.842% / 1.39 FPS、
seq07 0.514% / 1.39 FPS。fixed σ=0.4 (
--kc-lo-no-anneal) は RPE が同等か僅かに良く、 速度が 1.9-2.2x。KITTI + CV 予測では kernel-correlation が主効果で、annealing は convergence safety knob。 Artifacts:docs/benchmarks/kitti_full_new_methods/seq00_kc_lo_no_anneal.json,docs/benchmarks/kitti_full_new_methods/seq07_kc_lo_no_anneal.json,docs/benchmarks/kitti_full_new_methods/kc_lo_sigma_schedule_ablation.json。 - 完了: seq00/07 ベンチ JSON、README leaderboard 行 (M-GCLO と LODESTAR の間、seq07 トップ更新)、module README、methods.json (71手法)、memory 追記、commit。
papers/i_loam/ — Intensity Enhanced LOAM。出典: Yeong-Sang Park, Hyesu Jang, Ayoung Kim,
UR (Ubiquitous Robots) 2020。著者公開コード無し (RPM Lab repo に無く、GitHub の "I-LOAM"
ヒットは全て無関係な LOAM フォーク)。新規 web サーベイ (§00.6) の第一弾としてユーザが選定。
機構: LOAM の幾何 edge/plane パイプラインを保ちつつ、LiDAR 反射強度 (intensity) を scan-to-scan 対応付けに 2 経路で注入する。
- 強度拡張対応探索: 直線/平面を定義する 2/3 点目の候補コストを
‖Δp‖² + λ·ΔI²に して、幾何が同等でも反射強度が一致する候補を選ぶ (反復幾何・平行壁の曖昧性解消)。 - 強度類似度の残差重み: 採用した各対応の残差を
w = exp(−ΔI²/2σ²)で重み付け (新papers/i_loam/include/i_loam/intensity_factors.hの重み付き Ceres ファクタ)。 - Mapping (2026-06-09 追加): 強度強化 scan-to-scan の後、
aloam::LaserMappingで scan-to-map 精緻化 (A-LOAM/F-LOAM と同じ 3 段 LOAM パイプライン)。map 段は幾何のみ (強度は odometry 段のみ)。--i-loam-no-mappingで scan-to-scan ablation に分離。
重要な実装ポイント: 幾何特徴抽出は共有 papers/aloam の ScanRegistration を再利用するが、
aloam は PointXYZI の intensity フィールドを scan-id + rel_time のエンコードに転用する
ため、特徴抽出後に真の反射強度が失われる。よって I-LOAM は生入力点群への
pcl::KdTreeFLANN 最近傍参照で特徴点ごとに反射強度を復元する。
デフォルト intensity_sigma=0.15, intensity_corr_weight=1.0, enable_mapping=true。
- 統合済み: header/
intensity_factors.h/src/test (6/6 pass)、root + evaluation CMake、pcd_dogfooding(include / isSupportedMethod /ILoamDogfoodingOptionsstruct /runILoam(loadPCDXYZIで強度保持読み込み, leaf=0) / options var / arg-parse--i-loam-no-intensity/--i-loam-no-mapping/--i-loam-dense-profile/--i-loam-intensity-sigma/--i-loam-corr-weight/--i-loam-stride/ dispatch+print ブロック)、docs/methods.json(72手法, Intensity family)、module README、 seq00/07 ベンチ JSON (mapping ON + ablation OFF)。 - 結果 (KITTI, --no-gt-seed,
--i-loam-dense-profile, mapping ON (default)):- seq00: RPE 0.899% / ATE 12.7 m (~3 FPS)
- seq07: RPE 0.575% / ATE 1.7 m (~3 FPS)
- → from-paper leaderboard 同列 (A-LOAM ~0.61% seq07 水準)。README leaderboard 行追加済み。
- 強度 ablation (
--i-loam-no-mapping, scan-to-scan only): ON vs--i-loam-no-intensity:- seq00: 純幾何 3.186% → 2.606% = drift -18.2%
- seq07: 純幾何 3.806% → 3.055% = drift -19.7%
- mean intensity weight ≈0.76–0.79 (強度経路が実働)。
- raw artifacts:
docs/benchmarks/kitti_full_new_methods/seq00_i_loam_no_mapping.json,docs/benchmarks/kitti_full_new_methods/seq00_i_loam_no_intensity.json,docs/benchmarks/kitti_full_new_methods/seq07_i_loam_no_mapping.json,docs/benchmarks/kitti_full_new_methods/seq07_i_loam_no_intensity.json。 paired summary:docs/benchmarks/kitti_full_new_methods/i_loam_intensity_ablation.json。
- honest 所見: 反射強度が scan-to-scan で drift ~18-20% 改善 → 論文の中心主張が KITTI の未校正・粗い強度でも再現。mapping 統合で絶対 drift も competitive 水準に到達。 README には leaderboard 行 + ablation 節の両方を掲載 (ablation は scan-to-scan 分離)。
- 完了: seq00/07 ベンチ (mapping + ablation raw JSON)、README leaderboard + ablation 節、
module README、methods.json(72)、memory
i-loam-32nd-from-paper、PR #13 マージ (b02612c)、 mapping 追補 (2026-06-09)。
papers/pl_loam/ — LiDAR-monocular point+line visual odometry (ICRA 2020)。著者コード無し。
KITTI Odometry に RGB が無いため Velodyne→カメラ投影の深度グラディエント疑似画像上で
Harris コーナー + 勾配線分を抽出し、LiDAR パッチ中央値で深度、Ceres depth-prior 付き
point-line BA + スケール補正。pcd_dogfooding 統合、test_pl_loam 5/5 pass。
結果 (honest negative, pseudo-image): seq00 143.2% drift (ATE 3016 m)、seq07 116.9% (ATE 271 m).
KITTI Raw RGB 本評価 (2026-06-09): image_02 PNG + --pl-loam-rgb-root。
- raw_0009 200f: 99.6% drift, ATE 120 m, rgb_frames=200, mean_scale≈0.977
- raw_0061 200f: 99.3% drift, ATE 84 m, rgb_frames=200, mean_scale≈0.974
→ 実 RGB でも honest negative (~99% vs paper ~0.6–1%)。簡略 Harris/line front-end が主因。
Artifacts:
docs/benchmarks/kitti_full_new_methods/kitti_raw_*_200_rgb.json
papers/inten_loam/ — Intensity and Temporal Enhanced LOAM (RS 2022/23, arXiv:2209.05708)。
円筒 range+intensity 画像から ground/facade/edge/reflector 特徴を抽出し、幾何
(edge point-to-line, surf point-to-plane) + B-spline 強度登録 (8 ボクセル三線形の簡約版)
を Ceres で joint 最適化。
2026-06-09 拡張: TVF (temporal voxel filter) + DOR (dynamic object removal) +
scan-to-map mapping を追加。CLI: --inten-loam-no-{tvf,dor,mapping}。
test_inten_loam 5/5 pass、--inten-loam-no-intensity ablation あり。
結果 (honest negative):
Mapping ablation (2026-06-09, full seq00/07):
- Best drift: scan-to-scan (
--inten-loam-no-mapping) — seq00 52.7% / seq07 67.4% (matches original baseline). - Mapping lowers seq07 ATE (442→~300 m) but does not improve drift; on long seq00 mapping increases drift (52.7%→60.6%). Full pipeline (TVF+DOR+mapping) reaches seq00 68.4% / seq07 71.7% drift.
- Intensity registration remains near-neutral with mapping enabled.
- Artifacts:
seq{00,07}_inten_loam_ablation.json; scriptrun_inten_loam_ablation.py. - MulRan 120 cross-dataset: 185.4% drift (ATE 32 m) — KISS-ICP (225%) より drift は良いが ATE は悪化
- NCLT 600 cross-dataset: 46.5% drift (ATE 40 m) — intensity scan-to-map 群 (ICPSC 0.49%, MCC 0.47%) は良好だが InTEn は未転移 seq07 強度 OFF は 66.9% → 強度経路は near-neutral。
papers/nn_zupt/ — NN-ZUPT vehicle INS (Li et al., Meas. Sci. Technol. 2023,
doi:10.1088/1361-6501/acabde)。著者公開コード無し。機構: 50×6 IMU 窓の 1D-CNN
が停止確率を出力 → strapdown INS + ZUPT (速度ゼロ) + NHC。Python 学習 +
C++ 推論 (JSON weights)。pcd_dogfooding --methods nn_zupt (要 imu.csv)。
test_nn_zupt 3/3 pass、--nn-zupt-{threshold-only,no-zupt,no-nhc,prob} あり。
スコープ: 論文 Kalman 補正と前進速度 aiding は範囲外。
結果:
- KITTI OXTS 学習: 停止ラベルは drive_0011 中心 (~36%); val (stop-free drive) acc ~100%
- HDL-400 120f (cross-IMU): 92% drift / ATE 7.0 m / zupt_frames=0 (honest negative)
- NCLT 600f (cross-IMU): 27% drift / ATE 28 m / zupt_frames=0 完了: 学習スクリプト、weights JSON、ベンチ JSON、README、methods.json。
papers/nhc_net/ — NHC-Net adaptive non-holonomic constraints (Li et al., GPS
Solutions 2023)。著者公開コード無し。機構: 50×6 IMU 窓の VMSC CNN が運動状態
(停止/直進/旋回/スリップ) を分類し横/垂直速度を回帰 → strapdown INS に 適応 NHC
(信頼度スケール gain) + ZUPT (停止クラス)。Python 学習 + C++ 推論 (JSON weights)。
pcd_dogfooding --methods nhc_net (要 imu.csv)。test_nhc_net 3/3 pass。
スコープ: 論文 GNSS/INS EKF と前進速度 aiding (疑似オドメータ等) は範囲外;
横/垂直 NHC のみ (OdoNet と対比)。
結果:
- KITTI OXTS 学習: held-out drive val class acc ~100%, NHC MAE ~0.04 m/s
- HDL-400 120f (cross-IMU): 89% drift / ATE 6.7 m (前進速度無し → under-travel)
- NCLT 600f (cross-IMU): 32% drift / ATE 29 m (ZUPT 未発火 off-domain) 完了: 学習スクリプト、weights JSON、ベンチ JSON、README、methods.json (79手法)。
papers/odonet/ — OdoNet pseudo-odometer (arXiv:2109.03091 / IEEE Sensors J. 2022)。
著者公開コード無し。機構: 50×6 IMU 窓の 1D-CNN が前進速度を回帰 (scale 30 m/s)、
strapdown INS + NHC/ZUPT と融合。Python 学習 (train_odonet.py) + C++ 推論
(JSON weights)。pcd_dogfooding --methods odonet (要 imu.csv)。
test_odonet 4/4 pass、--odonet-{weights,no-nhc,no-zupt,nhc-only} あり。
結果:
- KITTI OXTS 学習: held-out drive val speed MAE ~2.1 m/s
- HDL-400 120f (cross-IMU): 890% drift / ATE 75 m (honest negative)
- NCLT 600f (cross-IMU): 249% drift / ATE 300 m (honest negative) 完了: 学習スクリプト、weights JSON、ベンチ JSON、README、methods.json (76手法)。
papers/vlom/ — Visual-LiDAR Odometry and Mapping with Monocular Scale Correction
and Visual Bootstrapping (arXiv:2304.08978, 2023)。著者公開コード無し。機構:
(1) 単眼 VO (PL-LOAM 流用の疑似画像特徴 + depth-prior BA)、(2) 三角測量深度と LiDAR
深度の比による MAD 外れ値除去付きスケール補正 (clamp 0.85–1.15)、(3) 視覚
frame-to-frame 運動を aloam::LaserOdometry::setMotionPrior で LiDAR 初期値に注入、
(4) A-LOAM scan-to-map mapping で最終ポーズ出力。
test_vlom 5/5 pass、--vlom-{fast,dense}-profile / --vlom-enable-bootstrap /
--vlom-no-bootstrap / --vlom-no-scale / --vlom-rgb-root あり。
結果:
- pseudo-image, bootstrap off default: seq00 0.91% / seq07 0.61% drift
- pseudo-image, old bootstrap-on path: seq00 89.2% / seq07 81.6% drift
- KITTI Raw RGB (2026-06-09): raw_0009 99.7% / raw_0061 99.3% drift (~99% both)
→ 実 RGB でも改善なし。mean_scale≈1.0, bootstrap≈199/200 frames。
完了: seq00/07 + Raw RGB ベンチ JSON、README、module README、
aloam::setMotionPriorAPI、methods.json (75手法)。
papers/icpsc/ — Intensity Cylindrical-Projection Shape Context (Zhang et al., JAG 2023)。
著者公開コード無し。機構: 円筒 range+intensity 画像から ICPSC 記述子 (ring×sector
平均強度 + log 密度) を構築、edge/surf/強度勾配特徴を抽出し、適応重み
w_geom = N_geom/(N_geom+α·N_int) で scan-to-map point-to-plane を融合。
ループクロージャ・因子グラフは 後回し (odometry ベンチに集中)。
test_icpsc 3/3 pass、--icpsc-{fast,dense}-profile / --icpsc-no-intensity あり。
結果 (mid-pack competitive): seq00 0.912% drift (ATE 19.2 m)、seq07 0.660%
(ATE 3.7 m, ~1.5 FPS)。両 seq で MCGICP-LO (0.940% / 0.774%) を上回る。
papers/mcgicp/ — Multi-Channel Generalized-ICP (Servos & Waslander, RAS 2017)。
著者公開コード無し。機構: LiDAR 強度を 4D 点共分散 (x,y,z,s·I) に統合し局所
Mahalanobis を推定、scan-to-map で強度重み付き point-to-plane を解く。
test_mcgicp 3/3 pass、--mcgicp-no-intensity ablation あり。
結果 (mid-pack competitive): seq00 0.940% drift (ATE 19.6 m)、seq07 0.774%
(ATE 4.5 m, ~2 FPS)。PLAN §00.6 の honest-negative 予測にはならず GMM-LO 帯。
papers/mcc_lo/ — Maximum Correntropy Criterion ロバスト point-to-plane オドメトリ。
出典: He et al., PLOS ONE 2018 (bidirectional MCC, doi:10.1371/journal.pone.0197542) +
Pattern Recognition 2019 (Correntropy scale ICP)。いずれも対点群レジストレーション論文で
オドメトリ公開実装は無い。機構: MSE の代わりに相関エントロピー V=Σexp(−r²/2σ²) を最大化、
半二次最適化で Welsch/Gauss 重み w=exp(−r²/2σ²)、カーネルバンド幅 σ を Silverman 経験則
σ=1.06·std(r)·N^(−1/5) で適応推定。注意点: 整備済みスキャンでは残差が小さく σ が
floor まで収縮し、狭すぎるカーネルが未整合点 (収束信号) まで棄却して 過小収束 する
(annealing 無し correntropy-ICP の既知の弱点)。そのため mcc_sigma_floor=0.3 を導入。
- 統合済み: header/src/test (3/3 pass)、root+evaluation CMake、
pcd_dogfooding8 スポット (include / method-check / DogfoodingOptions struct / runMccLo / options var / arg-parse--mcc-lo-{fast,dense}-profile/--mcc-lo-sigma/--mcc-lo-fixed-sigma/ dispatch / method-list 文字列 2 箇所)、docs/methods.json(68手法)、module README、seq07 ベンチ JSON。 - 結果: seq00 RPE 0.892% / ATE 12.9 m、seq07 RPE 0.611% / ATE 2.15 m, 2.6 FPS (mean_weight 0.90, σ=0.30 floor 張り付き) — robust 群最良 (GNC 0.986 / Student-T 0.952 / GMM 0.941 を上回り、seq00 ATE 12.9m も robust 群で最小)。seq07 0.611 は Adaptive-ICP 0.569 に次ぐ好成績。
- 完了: seq00/seq07 ベンチ JSON、README leaderboard 行 (Adaptive-ICP と NHC-LIO の間)、 module README、methods.json (68手法)、memory 追記、commit すべて完了。
3D 新規論文 shortlist は完了または保留。 アクティブ作業は §00.6c 2D キャンペーン。 以下は 3D カテゴリの記録。
純 LiDAR 幾何の distinct な古典機構は ほぼ枯渇 (saturated: ground/degeneracy/voxelmap/ feature-weighting/distribution-factor/motion-constraint/robust-M-estimator/robust-LTS(TrICP)/ spectral/gmm-em/correntropy/implicit-surface(IMLS)/correlation-entropy(KC))。そこでユーザ指示 「IMU dead-reckoning / LiDAR-visual / intensity-based LiDAR odometry を調査して再現実装して いこう」に基づき 3 カテゴリで新規 web サーベイを実施 (著者コード無し・KITTI 再現可能・ 既存非重複)。第一弾として I-LOAM(32) を完了 (§00.48)。残り shortlist (各カテゴリ ranked):
Intensity / reflectivity LiDAR
- ✅ I-LOAM (UR 2020) — 完了 (§00.48)。
- ✅ InTEn-LOAM (RS 2022/23) — 完了 (§00.50)。scan-to-scan ~53–67% drift; TVF/DOR/mapping 拡張済 (seq07 prefix 56.5% drift)。強度 ablation near-neutral。
- ✅ MCGICP (Servos & Waslander, RAS 2017) — 完了 (§00.51)。4D 共分散 + 強度重み付き point-to-plane。seq00 0.940% / seq07 0.774% drift (mid-pack competitive; honest-negative ではなく GMM-LO 帯)。
- ✅ ICPSC (JAG 2023) — 完了 (§00.52)。円筒 shape-context + 適応 geom/intensity 融合。 seq00 0.912% / seq07 0.660% (MCGICP 上回る mid-pack)。
LiDAR-visual odometry (KITTI は Velodyne + カメラ + GT 同期)
- ✅ PL-LOAM (ICRA 2020) — 完了 (§00.49)。Odometry 疑似画像 ~117–143% drift; KITTI Raw RGB 本評価完了 (~99% drift, honest negative)。
- ✅ VLOM (arXiv:2304.08978) — 完了 (§00.53)。KITTI Odometry 疑似画像は bootstrap off default で seq00 0.91% / seq07 0.61%; KITTI Raw RGB 本評価完了 (~99% drift, bootstrap-on honest negative)。
- (高難度・残) Tightly-coupled mono-LiDAR (HIT, Intell. Service Robotics 2022): 視覚再投影 + LiDAR 残差を単一最適化に密結合。paywall・要全文。feasibility 2.5/5。
- 除外 (著者/3rd-party コード有): LIMO, DEMO/Zhang, DVL-SLAM, SDV-LOAM, RGB-L。
IMU dead-reckoning / inertial odometry (KITTI 100Hz OXTS)
- ✅ OdoNet ⭐ (IEEE Sensors J. 2022, arXiv:2109.03091) — 完了 (§00.54)。1D-CNN 疑似オドメータ + INS/NHC/ZUPT。KITTI OXTS 学習 val speed MAE ~2.1 m/s; HDL-400/NCLT への cross-IMU 転移は honest negative (890% / 249% drift)。
- ✅ NHC-Net / VMSC (GPS Solutions 2023) — 完了 (§00.55)。CNN 運動状態分類 + 適応 NHC (横/垂直速度回帰)。KITTI OXTS 学習; HDL-400/NCLT cross-IMU honest negative (89% / 32% drift)。前進速度 aiding はスコープ外。
- ✅ NN-ZUPT (Meas. Sci. Technol. 2023) — 完了 (§00.56)。CNN 停止検出 + ZUPT/NHC。 KITTI 停止区間希薄; cross-IMU で zupt_frames=0 (honest negative)。
⚠️ このカテゴリの上位は 学習ベース が中心 → 純 C++ クラシカル campaign とワークフローが 異なる (Python 学習 + 重み移植)。着手前にユーザと方式合意が要る。- 除外 (コード有): Wheel-INS (i2Nav), AI-IMU Dead-Reckoning (Brossard)。
運用: 次 "tugi"/"tudukete" で 次の 1 候補を実装 (自律的に始めない)。IMU shortlist (OdoNet / NHC-Net / NN-ZUPT) は 完了。Intensity / LiDAR-visual shortlist は 完了 (RGB 本評価含む)。
- ✅ Cross-dataset 再評価 (2026-06-09): MulRan 120 + MulRan full (1177f) +
NCLT 600 完了 (
mulran_120_bundle.json,mulran_full_bundle.json,nclt_600_bundle.json)。MulRan full: InTEn-LOAM 59.1% drift (best); KISS/MCC ~103%; I-LOAM/MCGICP 発散。スクリプト:evaluation/scripts/run_cross_dataset_benchmark.py。README:docs/benchmarks/cross_dataset/README.md。 - ✅ validate_showcase + commit
8a980fcpush (2026-06-09) 完了。
3 カテゴリ shortlist 完了後の追加サーベイ。除外: OSS 済み (SiMpLE, MAD-ICP, RESPLE, MOLA, C3P-VoxelMap, VoxelMap++ 等)。
| Rank | 論文 | 機構 | OSS | Feasibility |
|---|---|---|---|---|
| ⭐1 | FR-LIO (arXiv:2302.04031) | sub-frame + ESKS + RC-Vox 固定配列 kNN マップ | 無 | 4/5 (要 IMU) |
| 2 | PG-LIO (2025) | NCC フォトメトリ + geom + IMU | 未公開 | 3.5/5 |
| 3 | ERASOR++ (2403.05019) | 動的物体除去 (静的マップ) | 無 | 3/5 (odom 本体ではない) |
| 4 | Tightly-coupled mono-LiDAR (ISR 2022) | 単眼+LiDAR 密結合 + LC | 無 | 2.5/5 |
推奨 41 本目: FR-LIO — RC-Vox は既存 voxelmap 群と非重複。NCLT/HDL-400/KITTI Raw
は imu.csv あり。LiDAR-only 簡約版も scoping 可。
運用: 「41本目 FR-LIO」等の明示で着手。
- ✅ 実装:
papers/fr_lio/— RC-Vox 二層ボクセルマップ、IMU 分散による adaptive sub-frame、簡略 backward ESKS、IMU 回転 preintegration + point-to-plane ICP。 - ✅ 統合: selector
fr_lio、--fr-lio-fast-profile/--fr-lio-dense-profile、imu.csv無しは skip（RKO-LIO と異なり IMU 必須）。 - ✅ テスト:
test_fr_lio4 cases PASS。 - ✅ NCLT 600 (
--no-gt-seed, fast profile, vs RKO-LIO/KISS-ICP):- FR-LIO: ATE 0.58 m, drift 0.63%
- RKO-LIO: ATE 1.13 m, drift 1.42%
- KISS-ICP: ATE 7.25 m, drift 2.76%
- 所見: 簡略 port だが NCLT では RKO/KISS を上回る。sub-frame は NCLT IMU では ほぼ常に 1（avg_subframes=1）— 動的区間では増加する設計。
- Artifact:
docs/benchmarks/cross_dataset/nclt_600_fr_lio_vs_rko.json - 未実装: deskew、full IEKF、重力/外参推定、著者コード無しのため論文数値との直接比較不可。
- 論文: Zhu, Zheng & Zhu, "Mesh-LOAM: Real-time Mesh-Based LiDAR Odometry and Mapping",
IEEE T-IV 2024, arXiv:2312.15630。公式 repo (
HelloXiaoZHU/Mesh-LOAM) は 2024 年から README のみでコード未公開、サードパーティ実装も無し。新規 web サーベイ (2026-06-11) の 第 1 推薦 (次点 ELO arXiv:2104.10879、他候補 RF-LIO / L-LO / LMBAO は PLAN 外メモ)。 - ✅ 実装:
papers/mesh_loam/— (1) passive-voxel IMLS-SDF 地図: 各点が 3×3×3 voxel 近傍に符号付き距離増分を scatter (O(N)、式 5–10、hybrid weightexp(-d²/h) + λ_n·max(0, nᵢ·n_v))、(2) dirty block 単位の増分 zero-surface 抽出、 (3) point-to-mesh GN odometry (CV 予測 + 曲率<0.1 平面特徴 + 0.2 m facet bin 探索 + |cos|>0.98 法線ゲート、式 3–4)。 - 主な逸脱 (詳細
papers/mesh_loam/README.md): marching cubes → marching tetrahedra (table-free)、GPU並列 hash/voxel deletion → CPUunordered_map+ radius prune、 2×2 m 可変高 block → 立方 8³ block、入力は harness で 0.25 m downsample (scatter 半径 0.3 m が隣接サンプルを覆い SDF 連結は維持)。 - ✅ テスト:
test_mesh_loam5 cases PASS (SDF+mesh 構築 / 静止 / 並進回復 / 系列追跡 / yaw 回復)。 - ✅ KITTI seq07 (108f window): ATE 0.135 m / RPE 0.525% (0.7 FPS)
- ✅ KITTI seq07 full (1101f,
--no-gt-seed --mesh-loam-dense-profile): ATE 0.98 m / RPE 0.616% / 0.005 deg/m / 0.74 FPS — 同 profile KISS-ICP (0.618%) と同等、from-paper 表 mid-top-10 圏。 Artifact:docs/benchmarks/kitti_full_new_methods/seq07_mesh_loam.json - ✅ KITTI seq00 full (4541f,
--no-gt-seed --mesh-loam-dense-profile): ATE 13.47 m / RPE 0.901% / 0.007 deg/m / 0.58 FPS — 同 profile KISS-ICP (0.872%) に seq00 RPE で僅差負け、from-paper 表は MCC-LO と NHC-LIO の間。 Artifact:docs/benchmarks/kitti_full_new_methods/seq00_mesh_loam.json再現コマンド:./build/evaluation/pcd_dogfooding data/kitti_pcd/seq00_full \ data/kitti_pcd/seq00_gt.csv --methods mesh_loam --no-gt-seed \ --mesh-loam-dense-profile \ --summary-json docs/benchmarks/kitti_full_new_methods/seq00_mesh_loam.json
- ✅ docs: README from-paper 表へ Mesh-LOAM 行追加、
papers/mesh_loam/README.mdの結果表更新。 - 状態: 実装 + harness + methods.json (92 手法) + seq00/07 full artifact + docs 更新済。
- 論文: Xin Zheng and Jianke Zhu, "Efficient LiDAR Odometry for Autonomous Driving", IEEE RA-L 2021, arXiv:2104.10879。著者公開実装は見つからず。論文/ KITTI detail は non-ground spherical range image + ground BEV map、range-adaptive normal、memory-efficient model update を主張。KITTI detail は SRI 2048×80、BEV 120 m × 60 m と記載。
- ✅ 実装:
papers/elo/— (1) non-ground SRI cell map (vertex + PCA normal)、 (2) ground BEV cell map (ground average + neighbor normal)、(3) projective frame-to-model point-to-plane GN、(4) registration 後に前 model を現フレームへ移し、stale cell を落として SRI/BEV を融合。 - 主な逸脱 (詳細
papers/elo/README.md): CUDA → CPU、full 実測は tractable runtime のため--elo-fast-profile(SRI 1024×80、BEV 0.3 m)。per-point time が無いので model window は frame index。ground segmentation は LiDAR 高 + 5 deg slope gate と欠損時 fallback。 - ✅ テスト:
test_elo5 cases PASS (初期 map / 静止 / 並進 / yaw / sequence)。 - ✅ KITTI seq00 full (4541f,
--no-gt-seed --elo-fast-profile): ATE 22.50 m / RPE 1.124% / 0.010 deg/m / 5.15 FPS — 論文 training seq00 0.54% には届かない honest negative だが、DiLO direct より大幅に安定。 Artifact:docs/benchmarks/kitti_full_new_methods/seq00_elo.json - ✅ KITTI seq07 full (1101f,
--no-gt-seed --elo-fast-profile): ATE 3.54 m / RPE 0.981% / 0.010 deg/m / 5.55 FPS — 論文 training seq07 0.31% には届かない。ground BEV は drift を抑えるが、CPU 簡略 port では KISS 同等までは届かず。 Artifact:docs/benchmarks/kitti_full_new_methods/seq07_elo.json - 再現コマンド:
./build/evaluation/pcd_dogfooding data/kitti_pcd/seq00_full \ data/kitti_pcd/seq00_gt.csv --methods elo --no-gt-seed \ --elo-fast-profile \ --summary-json docs/benchmarks/kitti_full_new_methods/seq00_elo.json
- ✅ docs: README from-paper 表へ ELO 行追加、
docs/methods.json93 手法、papers/elo/README.mdに seq00/07 結果表を追加。 - 状態: 実装 + harness + methods.json (93 手法) + seq00/07 full artifact + docs 更新済。
- 論文: Weizhuang Wu and Wanliang Wang, "LiDAR Inertial Odometry Based on Indexed Point and Delayed Removal Strategy in Highly Dynamic Environments", Sensors 2023 23(11):5188, doi:10.3390/s23115188。著者公開実装は見つからず。LIO-SAM ベースで pseudo occupancy による dynamic point detection、indexed point propagation、 delayed removal、dynamic weight を主張。
- ✅ 実装:
papers/id_lio/— (1) current scan / predicted map の spherical range image pseudo-occupancy、(2) map 点に stable id / confidence / missing age / dynamic age を保持する indexed voxel map、(3) delayed removal 前の低重み保持、 (4) source dynamic weight + map confidence 付き point-to-plane scan-to-map、 (5)imu.csvがある場合のみ gyro preintegration rotation prior。 - 主な逸脱 (詳細
papers/id_lio/README.md): 論文の LIO-SAM feature extraction / keyframe graph / GTSAM / loop closure / full IMU factor は範囲外。KITTI PCD export はframe_timestamps.csvのみでimu.csvが無いため、full KITTI は constant-velocity fallback。 - ✅ テスト:
test_id_lio5 cases PASS (初期 indexed map / 並進 sequence / IMU yaw prior / foreground dynamic downweight / delayed removal)。 - ✅ KITTI seq00 full (4541f,
--no-gt-seed --id-lio-dense-profile): ATE 15.45 m / RPE 1.111% / 0.014 deg/m / 7.82 FPS。 Dynamic machinery は active (dynamic/frame ≈464) だが、top KISS-like front-end には届かない honest mid-pack。Artifact:docs/benchmarks/kitti_full_new_methods/seq00_id_lio.json - ✅ KITTI seq07 full (1101f,
--no-gt-seed --id-lio-dense-profile): ATE 4.61 m / RPE 0.999% / 0.013 deg/m / 11.52 FPS。 Dynamic/frame ≈432。ELO と同程度だが seq07 は僅差で悪い。 Artifact:docs/benchmarks/kitti_full_new_methods/seq07_id_lio.json - 再現コマンド:
./build/evaluation/pcd_dogfooding data/kitti_pcd/seq00_full \ data/kitti_pcd/seq00_gt.csv --methods id_lio --no-gt-seed \ --id-lio-dense-profile \ --summary-json docs/benchmarks/kitti_full_new_methods/seq00_id_lio.json
- ✅ docs: README from-paper 表へ ID-LIO 行追加、
docs/methods.json94 手法、papers/id_lio/README.mdに seq00/07 結果表を追加。 - 状態: 実装 + harness + methods.json (94 手法) + seq00/07 full artifact + docs 更新済。
- 論文: Ke Cao et al., "Tightly-Coupled LiDAR-Visual SLAM Based on Geometric Features for Mobile Agents", ROBIO 2023 / arXiv:2307.07763。作者実装は見つからず。 LiDAR subsystem と monocular visual subsystem の geometric features を spherical fusion frame で対応づけ、visual line で LiDAR linear feature を補正し、LiDAR depth / direction で visual line landmark を補強する主張。
- ✅ 実装:
papers/tc_lvgf/— (1) spherical range-image fusion frame、 (2) range-image 局所 PCA による LiDAR linear feature、(3) KITTI PCD 用の pseudo-visual sparse range-line segment、(4) visual/LiDAR line direction fusion、 (5) point-to-plane + light point-to-line / direction residual scan-to-map、 (6) visual line 不足時の LiDAR fallback。 - 主な逸脱: RGB ORB/LSD / semantic association / dual SLAM backend /
loop closure は範囲外。KITTI Odometry PCD export は Velodyne のみなので、visual
line は LiDAR range-image から生成。短窓 ablation では強い line residual が僅かに悪化したため、
dense profile は
line_weight=0.2,direction_weight=0.05。 - ✅ テスト:
test_tc_lvgf4 cases PASS (球面投影 / line 抽出+fusion / 並進 tracking / visual sparse fallback)。 - ✅ KITTI seq00 full (4541f,
--no-gt-seed --tc-lvgf-dense-profile): ATE 11.95 m / RPE 1.055% / 0.011 deg/m / 8.43 FPS。 Artifact:docs/benchmarks/kitti_full_new_methods/seq00_tc_lvgf.json - ✅ KITTI seq07 full (1101f,
--no-gt-seed --tc-lvgf-dense-profile): ATE 3.74 m / RPE 0.941% / 0.011 deg/m / 11.21 FPS。 Artifact:docs/benchmarks/kitti_full_new_methods/seq07_tc_lvgf.json - 所見: pseudo-visual line は full で fallback 0 (約56 visual lines/frame) と安定。 PL-LOAM/VLOM の疑似画像 honest negative より大幅に良いが、最上位の KISS-like point-to-plane core はまだ超えない mid-pack positive。
- ✅ docs: README from-paper 表へ TC-LVGF 行追加、
docs/methods.json95 手法、papers/tc_lvgf/README.mdに seq00/07 結果表を追加。 状態: 実装 + harness + methods.json (95 手法) + seq00/07 full artifact + docs 更新済。
- 論文: Xuan He et al., "LiDAR-Visual-Inertial Odometry Based on Optimized Visual Point-Line Features", Remote Sensing 2022, 14(3):622。作者実装は web/GitHub survey で見つからず。improved visual point/line features、LiDAR depth association、 VIO-assisted LiDAR scan matching、Bayesian/factor-graph fusion、Helmert variance component weighting が主張。
- ✅ 実装:
papers/opl_lvio/— (1) KITTI PCD 用 range-image visual point proxy (high-curvature cells)、(2) smooth segment の pseudo visual line、 (3) LiDAR depth correlation 付き visual point/line map、(4) scan-to-map point-to-plane + visual point residual、(5) Helmert 風 residual variance weight adaptation、(6) visual feature 不足時の LiDAR fallback。 - 主な逸脱: RGB feature detector / IMU VIO / GNSS / loop closure / graph backend は
範囲外。KITTI Odometry PCD export は Velodyne のみなので visual point/line は LiDAR
range-image proxy。seq07 108f ablation では pseudo-line residual が僅かに悪化したため、
dense profile は
visual_point_weight=0.08,line_weight=0,direction_weight=0。 - ✅ テスト:
test_opl_lvio4 cases PASS (球面投影 / point+line 抽出 / point residual 付き並進 tracking / visual sparse fallback)。 - ✅ KITTI seq00 full (4541f,
--no-gt-seed --opl-lvio-dense-profile): ATE 15.24 m / RPE 1.050% / 0.011 deg/m / 7.67 FPS。 Artifact:docs/benchmarks/kitti_full_new_methods/seq00_opl_lvio.json - ✅ KITTI seq07 full (1101f,
--no-gt-seed --opl-lvio-dense-profile): ATE 3.65 m / RPE 0.902% / 0.011 deg/m / 11.18 FPS。 Artifact:docs/benchmarks/kitti_full_new_methods/seq07_opl_lvio.json - 所見: range-image visual points は fallback 0 (seq00 約94.8/frame、seq07 約98.8/frame) で安定。 RPE は TC-LVGF より僅かに良い (1.050/0.902 vs 1.055/0.941) が、seq00 ATE は悪い (15.24 m vs 11.95 m)。pseudo visual feature は有効だが、主成分は依然 point-to-plane core。
- ✅ docs: README from-paper 表へ OPL-LVIO 行追加、
docs/methods.json96 手法、papers/opl_lvio/README.mdに seq00/07 結果表を追加。 状態: 実装 + harness + methods.json (96 手法) + seq00/07 full artifact + docs 更新済。
00.52g V-LOAM2015 / TC-VLO / AD-VLO / TC-MVLO (97-100手法目 / 3D LiDAR-visual from-paper, 2026-06-12) — 実装 + seq00/07 full 完了
- 論文:
- V-LOAM2015: Ji Zhang and Sanjiv Singh, "Visual-Lidar Odometry and Mapping: Low-drift, Robust, and Fast", ICRA 2015。visual odometry bootstrap + LiDAR scan matching refinement。
- TC-VLO: Young-Woo Seo and Chih-Chung Chou, "A Tight Coupling of Vision-Lidar Measurements for an Effective Odometry", IEEE IV 2019。visual / LiDAR residual を pose solve に密結合。
- AD-VLO: Kaihong Huang, Junhao Xiao, Cyrill Stachniss, "Accurate Direct Visual-Laser Odometry with Explicit Occlusion Handling and Plane Detection", ICRA 2019。direct visual-laser alignment + occlusion / plane handling。
- TC-MVLO: Lingbo Meng, Chao Ye, Weiyang Lin, "A Tightly Coupled Monocular Visual Lidar Odometry with Loop Closure", Intelligent Service Robotics 2022。 monocular visual-LiDAR tight coupling + loop closure。
- ✅ 実装:
papers/v_loam15/,papers/tc_vlo/,papers/ad_vlo/,papers/tc_mvlo/。いずれもopl_lviobackend 上の paper-specific adapter: range-image visual point/line proxy、LiDAR depth付き visual residual、 scan-to-map point-to-plane refinement、Helmert-style residual scale balancingを共有し、 有効化する residual / weight / fallback 条件を論文ごとに分離。 - 主な profile 差分:
- V-LOAM2015: visual point bootstrap + small motion prior、line residual 無効。
- TC-VLO: visual point + point-to-line + direction residual を軽量に密結合。
- AD-VLO: range-jump occlusion gate + plane emphasis、line residual 無効。
- TC-MVLO: visual point + line residual を TC-VLO より強め、Helmert clamp を狭める。
- 主な逸脱: 4 本とも RGB camera tracker / true monocular VO / direct photometric image alignment / full SLAM graph / loop closure は範囲外。KITTI Odometry PCD export は Velodyne のみなので、visual features は LiDAR range-image proxy。PL-LOAM/VLOM の depth-gradient pseudo-image より安定するが、主成分は依然 point-to-plane core。
- ✅ テスト:
test_v_loam15,test_tc_vlo,test_ad_vlo,test_tc_mvlo各 3 cases PASS (profile 適用 / feature 抽出 / short translation tracking)。 - ✅ KITTI seq00 full (4541f,
--no-gt-seed, dense profile):- V-LOAM2015: ATE 14.37 m / RPE 1.066% / 7.76 FPS。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_v_loam15.json - TC-VLO: ATE 12.01 m / RPE 1.060% / 7.64 FPS。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_tc_vlo.json - AD-VLO: ATE 12.95 m / RPE 1.052% / 8.15 FPS。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_ad_vlo.json - TC-MVLO: ATE 10.82 m / RPE 1.054% / 8.00 FPS。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_tc_mvlo.json
- V-LOAM2015: ATE 14.37 m / RPE 1.066% / 7.76 FPS。
Artifact:
- ✅ KITTI seq07 full (1101f,
--no-gt-seed, dense profile):- V-LOAM2015: ATE 3.63 m / RPE 0.910% / 10.05 FPS。
- TC-VLO: ATE 4.22 m / RPE 0.925% / 10.91 FPS。
- AD-VLO: ATE 3.08 m / RPE 0.939% / 9.96 FPS。
- TC-MVLO: ATE 3.00 m / RPE 0.939% / 10.76 FPS。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq07_visual_lidar_97_100.json
- 所見: LiDAR-visual adapter family は 0.90-1.07% drift で安定し、PL-LOAM の pseudo-image honest negative より大幅に良い。VLOM も疑似画像 bootstrap を切れば 同じ帯に入る。AD-VLO は seq00 RPE、TC-MVLO は ATE、 V-LOAM2015 は seq07 RPE が強い。ただし KISS-like point-to-plane core を上回るほどではなく、 pseudo-visual residual は補助的。
- ✅ docs: README from-paper 表へ 4 行追加、
docs/methods.json100 手法、 各 module README と再現 artifact を追加。 状態: 実装 + harness + methods.json (100 手法) + seq00/07 full artifact + docs 更新済。
- 論文: Chenglong Qian et al., "RF-LIO: Removal-First Tightly-coupled Lidar Inertial Odometry in High Dynamic Environments", IROS 2021 / arXiv:2206.09463。 LIO-SAM 系の tightly coupled LIO に adaptive multi-resolution range image の removal-first dynamic object filtering を入れ、scan-match 前に moving foreground を落とす。
- ✅ 実装:
papers/rf_lio/— (1) 前フレーム static scan を予測 current sensor frame へ投影、 (2) fine/mid/coarse の multi-resolution range image、(3) 2 解像度以上で predicted surface より 近い点を foreground candidate として registration 前に除去、(4) removal cap / min keep guard、 (5)id_liobackend の indexed dynamic map / delayed removal / IMU gyro prior を再利用。 - 主な逸脱: full LIO-SAM factor graph / submap optimization / tightly coupled IMU backend は範囲外。
KITTI Odometry PCD は
imu.csvが無いので constant-velocity fallback。adaptive range image は compact previous-scan foreground test として実装。 - ✅ テスト:
test_rf_lio4 cases PASS (first map / foreground removal-first / short translation tracking / IMU prior pass-through)。 - ✅ KITTI seq07 short smoke (108f, dense): ATE 0.399 m / RPE 0.592% / 21.1 FPS。短窓では removal-first が良好。
- ✅ KITTI seq00 full (4541f,
--no-gt-seed --rf-lio-dense-profile): ATE 22.52 m / RPE 1.351% / 0.019 deg/m / 6.56 FPS。 Artifact:docs/benchmarks/kitti_full_new_methods/seq00_rf_lio.json - ✅ KITTI seq07 full (1101f,
--no-gt-seed --rf-lio-dense-profile): ATE 4.81 m / RPE 1.272% / 0.018 deg/m / 11.06 FPS。 Artifact:docs/benchmarks/kitti_full_new_methods/seq07_rf_lio.json - 所見: removal-first は active (seq00 273 点/frame、seq07 246 点/frame 除去) だが、 mostly static な KITTI では useful foreground も削り、ID-LIO (1.111/0.999%) より悪化。 paper の狙いは high dynamic scene なので、KITTI では stable-but-below-baseline の honest negative。
- ✅ docs: README from-paper 表へ RF-LIO 行追加、
docs/methods.json101 手法、papers/rf_lio/README.mdと seq00/07 artifact を追加。 状態: 実装 + harness + methods.json (101 手法) + seq00/07 full artifact + docs 更新済。
ユーザー指示: 「ronbun wo daseru kurai ni saigen zissou wo tyanto sasete, readme mo seibi sitai」。 新規手法追加を一段止め、論文に出せる再現実装 subset と広い catalog を切り分ける。
- ✅ 新規 docs:
docs/paper_ready_reproducibility.mdを追加。T0 paper-grade / T1 mechanism-grade / T2 adapter-grade / T3 smoke-concept の claim tier、paper-grade checklist、README claim policy を定義。 - ✅ README claim boundary: 101 手法の breadth catalog と、manuscript-level claim に使う
paper-ready subset を分離。
73 paper reimplementations/42 papers with no public author codeの表示を現在のdocs/methods.jsonと同期。 - ✅ reproducibility report:
docs/reproducibility_report.mdから paper-ready plan へ導線を追加。 - T0/T1 近傍候補:
- I-LOAM: intensity on/off の full artifact と paired commands を追加済み (§00.52j)。
- KC-LO: sigma schedule ablation と runtime/accuracy trade-off を追加済み (§00.52k)。
- M-GCLO: ground-factor off ablation は追加済み (§00.52l)。synthetic non-flat ground stress も追加済み (§00.52p)。 T0 昇格には public non-flat dataset validation が必要。
- Quadric-LO: plane-fallback ablation は追加済み (§00.52n)。synthetic curved-object stress も追加済み (§00.52q)。 T0 昇格には public curved-object / non-urban dataset validation が必要。
- RF-LIO/ID-LIO: synthetic dynamic-object stress は追加済み (§00.52o)。KITTI と synthetic だけで paper claim しない。public high-dynamic dataset validation を追加してから昇格判断。
- 次の既定動作: 新規 102 手法目ではなく、8-12 method の frozen paper table と paired ablation を
作る。README は breadth、論文本文は T0/T1 subset、T2/T3 は appendix catalog。
現在の seed bundle は
docs/benchmarks/paper_ready_bundle.json(§00.52m/§00.52n)。
paper-ready hardening の 1 本目として、I-LOAM の intensity on/off full-sequence paired artifacts を 生成し、既存 README の ablation 数値を raw JSON 付きにした。
- ✅ seq00 full, scan-to-scan only:
- intensity on: RPE 2.606375%, ATE 49.432759 m, FPS 8.52。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_i_loam_no_mapping.json - intensity off: RPE 3.185596%, ATE 76.028538 m, FPS 8.38。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_i_loam_no_intensity.json - delta: drift -18.18%, ATE -34.98%。
- intensity on: RPE 2.606375%, ATE 49.432759 m, FPS 8.52。
Artifact:
- ✅ seq07 full, scan-to-scan only:
- intensity on: RPE 3.055169%, ATE 15.132579 m, FPS 8.58。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq07_i_loam_no_mapping.json - intensity off: RPE 3.806092%, ATE 18.502306 m, FPS 8.88。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq07_i_loam_no_intensity.json - delta: drift -19.73%, ATE -18.21%。
- intensity on: RPE 3.055169%, ATE 15.132579 m, FPS 8.58。
Artifact:
- ✅ paired summary:
docs/benchmarks/kitti_full_new_methods/i_loam_intensity_ablation.json - ✅ docs: README ablation 節、
papers/i_loam/README.mdreproduce commands、docs/paper_ready_reproducibility.md、docs/reproducibility_report.mdを更新。 - 次: §00.52m の seed bundle に固定済み。次は Quadric-LO plane-fallback ablation、 RF-LIO/ID-LIO public high-dynamic dataset validation、または M-GCLO public non-flat dataset validation。
paper-ready hardening の 2 本目として、KC-LO の coarse-to-fine sigma annealing と fixed fine sigma を seq00/07 full で比較し、runtime/accuracy trade-off を raw artifact 付きにした。
- ✅ seq00 full:
- annealed σ 1.5→0.4: RPE 0.841902%, ATE 14.215286 m, FPS 1.39。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_kc_lo.json - fixed σ=0.4: RPE 0.837450%, ATE 13.404514 m, FPS 2.65。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_kc_lo_no_anneal.json - delta fixed vs annealed: RPE -0.53% relative, FPS +90.0%。
- annealed σ 1.5→0.4: RPE 0.841902%, ATE 14.215286 m, FPS 1.39。
Artifact:
- ✅ seq07 full:
- annealed σ 1.5→0.4: RPE 0.514327%, ATE 0.830931 m, FPS 1.39。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq07_kc_lo.json - fixed σ=0.4: RPE 0.509921%, ATE 0.858036 m, FPS 3.12。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq07_kc_lo_no_anneal.json - delta fixed vs annealed: RPE -0.86% relative, FPS +124.0%。
- annealed σ 1.5→0.4: RPE 0.514327%, ATE 0.830931 m, FPS 1.39。
Artifact:
- ✅ paired summary:
docs/benchmarks/kitti_full_new_methods/kc_lo_sigma_schedule_ablation.json - ✅ code metadata fix:
evaluation/src/pcd_dogfooding.cppの KC-LO note を修正し、 future artifact がfixed sigma (no annealing)/coarse-to-fine sigma annealingを正しく記録。 - ✅ docs: README KC-LO row を fixed-sigma best に更新、
papers/kc_lo/README.mdに ablation table と reproduce command 追加、paper-ready plan / reproducibility report 更新。 - 所見: KITTI full + constant-velocity prediction では、kernel-correlation 自体が主効果。 σ annealing は大域収束の保険で、well-initialized odometry では fixed fine σ が速くて同等精度。
- 次: §00.52m の seed bundle に固定済み。次は Quadric-LO plane-fallback ablation、 RF-LIO/ID-LIO public high-dynamic dataset validation、または M-GCLO public non-flat dataset validation。
paper-ready hardening の 3 本目として、M-GCLO の multiple-ground point-to-plane factor を
--m-gclo-no-ground で無効化し、paper mechanism の ground-on 既存 artifact と seq00/07 full で
比較した。ground off では分類された ground correspondence を統計だけ数え、最適化から外す。
- ✅ code:
evaluation/src/pcd_dogfooding.cppに--m-gclo-no-groundと note 分岐を追加。papers/m_gclo/src/m_gclo.cppはground_weight <= 0の ground correspondence を optimizer から除外する。
- ✅ seq00 full:
- ground on: RPE 0.835367%, rot 0.007618 deg/m, ATE 19.143798 m, FPS 3.41。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_m_gclo.json - ground off: RPE 0.833008%, rot 0.010510 deg/m, ATE 46.934202 m, FPS 5.85。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_m_gclo_no_ground.json - delta ground off vs on: translational RPE -0.28% relative, ATE +145.17%, rotational RPE +37.96%。
- ground on: RPE 0.835367%, rot 0.007618 deg/m, ATE 19.143798 m, FPS 3.41。
Artifact:
- ✅ seq07 full:
- ground on: RPE 0.670573%, rot 0.010557 deg/m, ATE 2.023877 m, FPS 3.67。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq07_m_gclo.json - ground off: RPE 0.600358%, rot 0.011158 deg/m, ATE 5.316227 m, FPS 5.68。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq07_m_gclo_no_ground.json - delta ground off vs on: translational RPE -10.47% relative, ATE +162.68%, rotational RPE +5.69%。
- ground on: RPE 0.670573%, rot 0.010557 deg/m, ATE 2.023877 m, FPS 3.67。
Artifact:
- ✅ paired summary:
docs/benchmarks/kitti_full_new_methods/m_gclo_ground_factor_ablation.json - ✅ docs: README M-GCLO paragraph、
papers/m_gclo/README.md、paper-ready plan、 reproducibility report を更新。 - 所見: KITTI seq00/07 では ground factor は translational RPE reducer というより global/attitude stability の anchoring term。RPE trans だけでは機構効果を過小評価する。
- 次: §00.52m の seed bundle に固定済み。次は Quadric-LO plane-fallback ablation、 RF-LIO/ID-LIO public high-dynamic dataset validation、または M-GCLO public non-flat dataset validation。
paper-ready hardening の 4 本目として、I-LOAM / KC-LO / M-GCLO の raw artifacts と paired ablations を 1 つの frozen evidence manifest に固定した。
- ✅ script:
evaluation/scripts/build_paper_ready_bundle.py- committed raw JSON と paired summary を読み、summary metric と raw artifact metric の不一致を検出。
--checkで manifest の stale check が可能。- reproduce command はすべて repo 相対 path (
data/kitti_pcd/...,docs/benchmarks/...) で記録。
- ✅ manifest:
docs/benchmarks/paper_ready_bundle.jsonbundle_id:paper_ready_core_2026_06_12- frozen methods: 3 (後続 §00.52n で 4 に拡張)
- paper table rows: 6 (後続 §00.52n で 8 に拡張)
- paired ablation summaries: 3 (後続 §00.52n で 4 に拡張)
- T0 evidence candidates: I-LOAM, KC-LO
- T1+ evidence candidates: M-GCLO
- ✅ docs: README claim boundary、
docs/paper_ready_reproducibility.md,docs/reproducibility_report.md, PLAN を bundle に同期。 - ✅ validation:
python3 evaluation/scripts/build_paper_ready_bundle.py --checkpython3 -m json.tool docs/benchmarks/paper_ready_bundle.json
- 次: Quadric-LO plane-fallback ablation、RF-LIO/ID-LIO public high-dynamic dataset validation、 または M-GCLO public non-flat dataset validation で seed bundle を 8-12 method table に拡張。
paper-ready hardening の 5 本目として、Quadric-LO の point-to-plane fallback を
--quadric-lo-no-plane-fallback で無効化し、paper mechanism の fallback-on 既存 artifact と
seq00/07 full で比較した。fallback off では quadric が立たない対応を optimizer から外す。
- ✅ code:
papers/quadric_lo/include/quadric_lo/quadric_lo.hにallow_plane_fallbackを追加。papers/quadric_lo/src/quadric_lo.cppで fallback 分岐を無効化可能にした。evaluation/src/pcd_dogfooding.cppに--quadric-lo-no-plane-fallbackとplane_fallback_rationote を追加。
- ✅ seq00 full:
- fallback on: RPE 0.867316%, rot 0.007640 deg/m, ATE 14.736361 m,
FPS 0.62, fallback ratio 0.005757。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_quadric_lo.json - fallback off: RPE 0.879792%, rot 0.007823 deg/m, ATE 12.894692 m,
FPS 1.13, fallback ratio 0.0。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq00_quadric_lo_no_plane_fallback.json - delta off vs on: translational RPE +1.44% relative, ATE -12.50%, FPS +83.34%。
- fallback on: RPE 0.867316%, rot 0.007640 deg/m, ATE 14.736361 m,
FPS 0.62, fallback ratio 0.005757。
Artifact:
- ✅ seq07 full:
- fallback on: RPE 0.597674%, rot 0.006249 deg/m, ATE 2.129764 m,
FPS 0.67, fallback ratio 0.005078。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq07_quadric_lo.json - fallback off: RPE 0.589873%, rot 0.006165 deg/m, ATE 2.146722 m,
FPS 1.06, fallback ratio 0.0。
Artifact:
docs/benchmarks/kitti_full_new_methods/seq07_quadric_lo_no_plane_fallback.json - delta off vs on: translational RPE -1.31% relative, ATE +0.80%, FPS +58.73%。
- fallback on: RPE 0.597674%, rot 0.006249 deg/m, ATE 2.129764 m,
FPS 0.67, fallback ratio 0.005078。
Artifact:
- ✅ paired summary:
docs/benchmarks/kitti_full_new_methods/quadric_lo_plane_fallback_ablation.json - ✅ bundle:
docs/benchmarks/paper_ready_bundle.jsonを 4 methods / 8 rows / 4 ablations に拡張。 - ✅ docs: README Quadric paragraph、
papers/quadric_lo/README.md、paper-ready plan、 reproducibility report、PLAN を更新。 - 所見: KITTI seq00/07 では plane fallback は used correspondence の約 0.5-0.6% と少なく、 Quadric-LO の evidence は fallback ではなく point-to-quadric path が支配的。T0 昇格には public curved-object / non-urban dataset validation が残る。
- 次: RF-LIO/ID-LIO public high-dynamic dataset validation、M-GCLO public non-flat dataset validation、 または Quadric-LO public curved-object / non-urban dataset validation。
paper-ready hardening の 6 本目として、RF-LIO/ID-LIO の dynamic filtering を KITTI 以外で
機構確認する dataset-free stress を追加した。自車は直進のみ、静的 corridor + 固定 landmark に
横断する box-shaped foreground objects を混ぜる。clean fixture と dynamic fixture を同一 runner で
生成し、30-frame window を pcd_dogfooding に流す。
- ✅ generator:
evaluation/scripts/generate_dynamic_object_stress_fixture.pyevaluation/fixtures/dynamic_object_stress_cleanevaluation/fixtures/dynamic_object_stress- fixture は生成物なので
.gitignoreで除外。再現は runner から行う。
- ✅ runner:
evaluation/scripts/run_dynamic_object_stress.py- clean/default:
docs/benchmarks/dynamic_object_stress/static_clean_default_30.json - dynamic/default:
docs/benchmarks/dynamic_object_stress/dynamic_objects_default_30.json - dynamic/RF conservative:
docs/benchmarks/dynamic_object_stress/dynamic_objects_rf_lio_conservative_30.json - summary:
docs/benchmarks/dynamic_object_stress/rf_id_lio_dynamic_object_stress_summary.json
- clean/default:
- ✅ 結果 (30 frames, no GT seed):
- KISS sanity ref: clean ATE 0.028915 m → dynamic 0.040450 m。
- ID-LIO: clean ATE 0.675926 m → dynamic 130.549154 m、
dynamic/frame=78.566667。 - RF-LIO default: clean ATE 2.486800 m → dynamic 49.932244 m、
removal_first/frame=214.266667。 - RF-LIO conservative (
--rf-lio-foreground-margin 1.5 --rf-lio-max-removal-fraction 0.08): dynamic ATE 41.632002 m、default RF 比 ATE -16.62%。
- 所見: synthetic foreground boxes は両 dynamic path を強く activate する。RF-LIO は ID-LIO より failure severity を抑え、removal cap も効く。ただし KISS sanity reference はほぼ崩れないため、 これは mechanism/failure-boundary artifact であって、paper-grade dynamic-scene claim には public high-dynamic dataset validation が必要。
- docs: README、
papers/id_lio/README.md、papers/rf_lio/README.md、docs/paper_ready_reproducibility.md、docs/reproducibility_report.md、paper-ready bundle status を更新。 - 次: M-GCLO public non-flat dataset validation、Quadric-LO public curved-object / non-urban dataset validation、 または RF-LIO/ID-LIO public high-dynamic dataset validation。
paper-ready hardening の 7 本目として、M-GCLO の multiple-ground factor を KITTI 以外の rolling-ground synthetic fixture で stress した。ローカルの prepared NCLT/MulRan PCD はこの worktree には残っていなかったため、今回の artifact は public dataset validation ではなく dataset-free mechanism stress として扱う。
- ✅ generator:
evaluation/scripts/generate_nonflat_ground_stress_fixture.py- rolling ground + low banks/curbs + non-ground landmark を生成。
- fixture は
evaluation/fixtures/nonflat_ground_stressに生成し、.gitignoreで除外。
- ✅ runner:
evaluation/scripts/run_m_gclo_nonflat_stress.py- ground on + KISS sanity:
docs/benchmarks/nonflat_ground_stress/m_gclo_nonflat_ground_on_60.json - ground off:
docs/benchmarks/nonflat_ground_stress/m_gclo_nonflat_ground_off_60.json - summary:
docs/benchmarks/nonflat_ground_stress/m_gclo_nonflat_ground_stress_summary.json
- ground on + KISS sanity:
- ✅ 結果 (60 frames, no GT seed):
- KISS sanity/failure ref: ATE 22.159251 m, RPE 180.223576%。
- M-GCLO ground on: ATE 0.116132 m, RPE 0.499712%,
rot 0.002823 deg/m,
mean_ground/nonground_corr=1278/241。 - M-GCLO ground off: ATE 0.149619 m, RPE 0.674768%,
rot 0.020621 deg/m,
mean_ground/nonground_corr=1288/231。 - delta ground off vs on: ATE +28.84%, RPE +35.03%, rotational RPE +630%。
- 所見: rolling-ground stress では multiple-ground path が明確に効く。KITTI ablation では ground off が translational RPE だけなら同等/低めだったが、non-flat synthetic では ATE/RPE/rot が すべて悪化する。T0 昇格には public non-flat dataset validation がまだ必要。
- docs: README、
papers/m_gclo/README.md、paper-ready plan、 reproducibility report、paper-ready bundle status を更新。 - 次: Quadric-LO public curved-object / non-urban dataset validation、RF-LIO/ID-LIO public high-dynamic dataset validation、 または M-GCLO public non-flat dataset validation。
paper-ready hardening の 8 本目として、Quadric-LO の point-to-quadric path を orchard-like な curved-object / non-urban synthetic fixture で stress した。木の幹 (cylinders)、岩/低木 (ellipsoids)、疎な起伏地面で構成し、平面 fallback ではなく quadric 対応が支配的になるようにした。
- ✅ generator:
evaluation/scripts/generate_quadric_curved_stress_fixture.pyevaluation/fixtures/quadric_curved_stressに生成。- fixture は生成物なので
.gitignoreで除外。再現は runner から行う。
- ✅ runner:
evaluation/scripts/run_quadric_curved_stress.py- fallback on + KISS sanity:
docs/benchmarks/quadric_curved_stress/quadric_curved_fallback_on_60.json - fallback off:
docs/benchmarks/quadric_curved_stress/quadric_curved_fallback_off_60.json - summary:
docs/benchmarks/quadric_curved_stress/quadric_curved_stress_summary.json
- fallback on + KISS sanity:
- ✅ 結果 (60 frames, no GT seed):
- KISS sanity ref: ATE 0.018757 m, RPE 0.118534%。
- Quadric-LO fallback on: ATE 0.068929 m, RPE 0.611652%,
rot 0.013907 deg/m,
mean_quadric/plane_corr=1242/12,plane_fallback_ratio=0.0094。 - Quadric-LO fallback off: ATE 0.069924 m, RPE 0.608845%,
rot 0.015174 deg/m,
mean_quadric/plane_corr=1245/0。 - delta fallback off vs on: ATE +1.44%, RPE -0.46%, rot +9.11%。
- 所見: curved synthetic では対応の 99% 以上が point-to-quadric。plane fallback を切っても ATE/RPE はほぼ不変で、KITTI ablation と同じく evidence は fallback ではなく quadric path が支配的。 ただし KISS sanity reference も非常に良く、これは public curved/non-urban dataset validation の代替ではない。
- docs: README、
papers/quadric_lo/README.md、paper-ready plan、 reproducibility report、paper-ready bundle status を更新。 - 次: RF-LIO/ID-LIO public high-dynamic dataset validation、M-GCLO public non-flat dataset validation、 または Quadric-LO public curved-object / non-urban dataset validation。
- ✅ 実装:
papers/pg_lio/— range-image NCC パッチ + point-to-plane ICP + 幾何退化時 photometric 重み増加 + IMU 並進 prior。 - ✅ 統合: selector
pg_lio、--pg-lio-fast-profile/--pg-lio-dense-profile、imu.csv+ PointXYZI 必須（欠如時 skip）。 - ✅ テスト:
test_pg_lio4 cases PASS。 - ✅ NCLT 600 (
--no-gt-seed, fast profile, vs FR-LIO):- FR-LIO: ATE 0.58 m, drift 0.63%
- PG-LIO (初版): ATE 23.7 m, drift 33.0% — ** honest negative**
- 所見: 簡略 NCC + 非 organized Velodyne 投影で photometric 因子が不安定。 patch 上限・退化時のみ photometric 適用・IMU 並進 prior 除去で改善余地あり。
- Artifact:
docs/benchmarks/cross_dataset/nclt_600_pg_lio_vs_fr_lio.json - 未実装: sliding-window factor graph、deskew、Ouster bias LUT、full IEKF。
状態: ワークツリーに未コミット。NCLT 600 で honest negative のため、当面は 2D LiDAR 路線を優先（下記 §00.6c）。
ユーザー指示: 「しばらくは 2D LiDAR でやっていく」。3D Velodyne + IMU 系 (FR-LIO 完了、PG-LIO は保留) から、planar laser scan 系 odometry/SLAM へ軸足を移す。
| Phase | 内容 | 状態 |
|---|---|---|
| Phase 1a | harness + RF2O/PL-ICP/CSM/Kinematic-ICP | ✅ push 済 |
| Phase 1b | 評価インフラ (drift, bag prep, corridor) | ✅ push 済 |
| Phase 1c | PSM + Intel CI smoke | ✅ push 済 (709d25c) |
| Phase 1d | Bonn 公開 fixtures (intel/fr079/mit) + 5法ベンチ | ✅ push 済 (a2817ff) |
| Phase 1e | NDT-2D (48本目) | ✅ push 済 (8ea40f1) |
| Phase 1f | IDC (49本目) | ✅ commit、未 push (361a592) |
| Phase 1g | CSM DT + pyramid refresh | ✅ commit、未 push (3fc5be0) |
| Phase 1h | ドキュメント整備 (scan2d hub) | ✅ ワークツリー、未 commit |
| Phase 2a | MbICP (50本目) + 8-method refresh | ✅ ワークツリー |
- canonical JSON 統合:
run_scan2d_benchmark.shで 8 法を 1 JSON/fixture に refresh 済み。 - MIT val: 33 frame のみ — drift は indicative。長 window が必要。
- 合成 corridor: MbICP 0.46% が PL-ICP 0.38% に近接。CSM は ~73% で honest negative 継続。
- local map / SLAM graph: 全 8 法 scan-to-scan のみ — Karto/Hector 級には未達。
- ROS2 / 実 bag CI: smoke は Intel 20f slice のみ。full bag export は手動。
| 要素 | 状態 |
|---|---|
| Harness | ✅ evaluation/src/scan_dogfooding.cpp |
| 入力 | ✅ scan_meta.json + NNNNNNNN/scan.csv |
| GT | ✅ gt.csv: timestamp, x, y, yaw [rad] |
| 評価 | ✅ 2D ATE + drift [%] (KITTI-style RPE segment 規則) + --summary-json |
| GT-seed | ✅ frame 0 anchor; --no-gt-seed 対応 |
| 前処理 (ROS1) | ✅ prepare_2d_scan_inputs.py |
| 前処理 (Bonn JSON) | ✅ prepare_bonn_2dslam_inputs.py |
| 合成 fixture 生成 | ✅ generate_rf2o_smoke_fixture.py, generate_rf2o_corridor_fixture.py |
| GT 可視化 | ✅ plot_scan2d_gt_overview.py → docs/assets/scan2d_public_gt.png |
| セットアップ doc | ✅ evaluation/scripts/SETUP_2D_SCAN_BENCHMARK.md |
| 一括ベンチ | ✅ evaluation/scripts/run_scan2d_benchmark.sh |
| CI smoke | ✅ evaluation/scripts/smoke_scan2d_fixture.sh (Intel 20f × 8 methods) |
| 合成ベンチ | ✅ rf2o_smoke (60f) + rf2o_corridor (120f, slow motion) |
| 公開 log | ✅ Bonn 2D-SLAM JSON → intel_val_73, fr079_val_384, mit_val_33 |
| リーダーボード | ✅ docs/benchmarks/scan2d/README.md |
| サマリ JSON | ✅ docs/benchmarks/scan2d/public_bundle.json |
| Rank | 論文 | Paper # | 状態 |
|---|---|---|---|
| ⭐1 | RF2O (ICRA 2016) | 43 | ✅ |
| 2 | PL-ICP (IROS 2008) | 44 | ✅ |
| 3 | Kinematic-ICP 2D (ICRA 2025) | 46 | ✅ |
| 4 | CSM / Karto (ICRA 2009) | 45 | ✅ (+ DT refresh) |
| 5 | PSM (ICRA 2003) | 47 | ✅ |
| 6 | NDT-2D (IROS 2003) | 48 | ✅ |
| 7 | IDC (Lu & Milios 1997) | 49 | ✅ |
| Rank | 論文 | 機構 | OSS | Feasibility | 備考 |
|---|---|---|---|---|---|
| ✅ | MbICP (Minguez et al., ICRA 2005 / T-RO 2006) | config-space metric ICP、対称性 | 各所 | 完了 | 50本目 |
| 2 | PL-ICP + local map | Censi local map / CSM 併用 | — | 5/5 | ✅ robot-frame cache + harness (Intel 15.0%, fr079 14.1%) |
| 3 | Hector SLAM scan matcher | Gauss-Newton on occupancy grid | ROS 有 | 3.5/5 | OSS あり → campaign 対象外の可能性 |
| 4 | Olson 2009 full Karto | 確率 grid + branch-and-bound | OpenSLAM | 3/5 | CSM からの段階的拡張 |
| 5 | GMapping particle filter | Rao-Blackwellized PF | ROS 有 | 2.5/5 | SLAM 本体、odom 単体ではない |
次の推奨: PG-LIO (3D) は引き続き保留。2D 側は新 fixture 追加より Felzenszwalb 後の drift 安定性 watch。
- ✅ 実装:
papers/rf2o/— range-flow 密拘束 + Gaussian pyramid + coarse-to-fine IRLS + 共分散 eigenvalue smoothing（MAPIRlab RF2O 準拠のループ順・PoseUpdate）。 - ✅ Harness 初版:
scan_dogfooding—scan_meta.json+NNNNNNNN/scan.csv+gt.csv。 - ✅ テスト:
test_rf2o8 cases PASS。 - ✅ Public benchmark (Bonn intel val, GT-seed, 73f, 378 m):
- Drift 14.3% — Intel リーダー
- fr079 15.4% / MIT 27.6% / corridor 1.3%
- ✅ Smoke (synthetic 60f): ATE 0.20 m, drift ~1.1%
- ✅ Robot-frame local map 実装 (opt-in, default off) (2026-06-10, P18) — see §00.6c-43 RF2O local map
- 未実装: Cauchy IRLS 再重み付け
- 状態: ✅ push 済; P18 local map は honest negative as default で opt-in 維持
- ✅ 実装:
papers/pl_icp/— beam-order line normals + iterative PL-ICP Gauss-Newton (SE2)。 - ✅ 統合:
scan_dogfooding --methods pl_icp（motion prior warm-start）。 - ✅ テスト:
test_pl_icp5 cases PASS。 - ✅ Public benchmark:
- Intel 16.9% / fr079 41.0% / MIT 30.3%
- Corridor 0.4% — 合成 fixture リーダー (RF2O 1.3% に次点)
- ✅ 所見: scan-to-scan ICP は長軌道 public log でドリフト蓄積。短 corridor では最強。
- 未実装: Censi recursive distortion、local map、kd-tree
- 状態: ✅ push 済
初版 (commit 5e2147c):
- brute-force SE(2) + Gaussian endpoint score + motion prior warm-start
- smoke 60f: drift ~19% vs RF2O ~1.1%
DT + pyramid refresh (commit 3fc5be0):
- ✅ occupancy grid + chamfer distance transform scoring (Olson-style)
- ✅ 3-level resolution pyramid (coarse → fine)
- ✅ relative SE(2) search around motion prior + bilinear DT lookup
- ✅ Public benchmark (after refresh):
- Intel 16.0% (was 17.6%) / fr079 20.6% (was 38.9%) / MIT 27.7%
- Corridor 73.3% (was 69.6% → 初版 DT で 167% 退行後 fix)
- ✅ 所見 (honest): DT は real Bonn logs で大幅改善 (特に fr079) だが 合成 corridor では依然 ~73% — PL-ICP 0.4% / RF2O 1.3% に大差。
Robot-frame local map (2026-06-10, commit 63576de):
- ✅
use_local_map+ voxel merge + radius prune + adaptive search (Karto/MbICP パターン) - ✅ harness 有効化 (
runCSM) - ✅ Public benchmark (local map):
- Intel 14.5% (was 16.0%) / fr079 14.5% (was 20.6%) / corridor 95.8% (was 73.3%)
Olson coarse branch-and-bound (2026-06-10, pushed):
- ✅ score upper-bound pyramid + coarse BnB + fine refinement (Karto から移植)
- ✅ harness デフォルト
use_branch_and_bound=true - ✅ Public benchmark (BnB refresh):
- Intel 15.2% / fr079 14.9% / corridor 102.0%
- ✅ 所見 (honest): BnB は探索スタックを Karto 級に揃えるが、Bonn drift は local map 単体から微悪化〜横ばい。corridor honest negative 継続。
速度チューニング (2026-06-10, P9):
- ✅ BnB node budget 128→64、leaf 3→2、fine refine 5→3、中間 pyramid refine スキップ
- ✅ score grid bilinear lookup（
exp()削減） - ✅ Public benchmark (tuned):
- Intel 14.7% (~79 FPS, was ~16 FPS) / fr079 14.3% (~58 FPS, was ~11 FPS) / corridor 41.3% (was 102%)
- ✅ 所見: 速度最適化が Bonn drift を改善し corridor も大幅改善。Felzenszwalb EDT は次候補 → P11 完了。
Felzenszwalb EDT (2026-06-10, P11):
- ✅
common/felzenszwalb_edt— exact Euclidean DT replaces chamfer in CSM + Karto - ✅ Public benchmark (Felzenszwalb EDT):
- Intel 14.0% (was 14.7%) / fr079 13.7% (was 14.3%) / corridor 30.5% (was 41%)
- ✅ 所見 (honest): main Bonn val + corridor improve; short
fr079_train_200regressed 12%→40% (indicative window). - 状態: 実装済
- ✅ 実装:
papers/kinematic_icp/— PL-ICP + wheel odom 重み付き prior + unicycle 射影。 - ✅ 統合:
--methods kinematic_icp --wheel-odom-from-gt - ✅ テスト:
test_kinematic_icp7 cases PASS。 - ✅ Public benchmark:
- Intel 18.4% / fr079 18.9% / MIT 23.4% (MIT リーダー) / corridor 83.8%
- ✅ 所見: GT wheel odom 合成でも smoke 高速運動では RF2O に劣後 (honest negative)。 短 MIT window では best drift だが 33f は indicative のみ。
- ✅ Robot-frame local map 実装 (opt-in, default off) (2026-06-11, P19) — see §00.6c-46 Kinematic-ICP local map
- 未実装: PRBonn 3D pipeline、動的重み付け、extrinsic TF
- 状態: ✅ push 済
- ✅ Point 型 robot-frame rolling local map — PL-ICP/MbICP と同構造 (RefPoint voxel merge + radius prune + grid-indexed correspondence + normal 回転)。
solveKinematicIncrementIndexed+ 共通finalizeKinematicSolve(unicycle prior + 射影は共有) - ✅ テスト:
test_kinematic_icp7 cases PASS (+LocalMapTracksTranslationWithWheelOdom,LocalMapMatchesScanToScanOnShortRun) - ✅ 全 fixture sweep (radius 10–20 m × voxel 0.10–0.25 m × wheel weight 8–64):
- 劇的改善:
mit_train_12046.4%→12.8% /fr079_train_20011.0%→5.9% / fr079 val 18.9%→16.0% - 同時に破壊: MIT val 23.4%→30.5% (唯一のリーダー fixture) /
fr079_train_120010.7%→18.7% / corridor 83.8%→93.5% - wheel weight 増 (correspondence 数増の補償) は train_200 を 5.9%→81.8% に不安定化
- 劇的改善:
- ✅ 判断: opt-in (default off) — RF2O P18 (projection 型は long 窓で全敗) と違い、point 型 map は真の trade-off だが dominating config が無い。「fixture 依存が支配的」の追加証拠。scan-to-scan (paper-faithful) を default 維持
- ✅ 詳細:
papers/kinematic_icp/README.md"Local map: opt-in only, no safe default" 節 - 状態: ✅ 完了 (canonical JSON 変更なし — default 不変)
- ✅ 実装:
papers/psm/— polar range-profile Gaussian correlation + coarse-to-fine search。 - ✅ 統合:
scan_dogfooding --methods psm - ✅ テスト:
test_psm7 cases PASS。 - ✅ Intel CI smoke 導入 (
709d25c):smoke_scan2d_fixture.sh - ✅ Public benchmark (local map, 2026-06-10):
- Intel 15.3% / fr079 14.3% / MIT 28.5% / corridor 4.4%
- ✅ 所見 (honest): local map で Intel/corridor 大幅改善。fr079 val は mid-pack (~Karto 13.7%)。long train は NDT/PL-ICP 劣後。 co-occurrence matrix 無しの簡略 polar correlation。
- 未実装: occupancy/co-occurrence matrix
- 状態: ✅ 実装済 (P16 local map)
- ✅ Robot-frame rolling local map 実装 — voxel merge + radius prune + per-point max-age expiry; accumulated points → min-range polar profile as range-flow reference (
setReferenceFromRanges) - ✅ テスト:
test_rf2o8 cases PASS (+RobotFrameLocalMapTracksTranslation,LocalMapImprovesLongTranslationVsScanToScan,LocalMapMaxAgeExpiresStalePoints) - ✅ 全 fixture sweep 完了 (radius 8–20 m × voxel 0.05–0.25 m × max-age 2–50f, 13 configs × 8 fixtures):
- val 改善は再現: fr079 val 13.9% (was 15.4%) / corridor 0.6% (was 1.3%) / Intel・MIT ~flat
- long train は全 config で退行:
fr079_train_120010.6%→best 20.1% /fr079_train_20030.3%→best 87.8% /intel_train_15017.5%→21.8% - age 減衰でも救えない (age2 でも fr079_train_200 ~93%) → stale/動的点は根本原因ではなく、min-range polar 投影自体 (bin 量子化 + 近距離バイアス) が range-flow 勾配を壊す
- IDC 85% / PSM 681% (fr079_train_200) と同族パターン = projection-style local map の family-level finding
- ✅ 判断: default OFF (opt-in) — scan-to-scan が既に Intel val リーダーで val 改善は微小、long 窓の損失が支配的。CT-ICP coarse-to-fine の opt-in 前例と整合。
RF2OParams::use_local_map+ 3 テストは維持 - ✅ canonical JSON を scan-to-scan default で全 refresh (
run_scan2d_benchmark.sh+run_scan2d_long_benchmark.sh、fr079_train_1200含む) +public_bundle.json再生成 - ✅ 詳細:
papers/rf2o/README.md"Local map: honest negative as a default" 節 - 未実装: Cauchy IRLS 再重み付け
- 状態: ✅ 完了
次の推奨: PG-LIO (3D) は引き続き保留。2D local map 化は P19 で全 9 法調査完了。残りは docs polish (methods.json 2D tag / validate_showcase 2D セクション)、または 3D 側の新論文。
- ✅ Outlier trimming — range-jump filter on match points (not local map); Mahalanobis trim (90%) on finest pyramid level only
- ✅ harness 有効化 (
use_outlier_trimming=true,max_range_jump=0.5,trim_fraction=0.9) - ✅ Public benchmark:
- Intel 14.8% (was 14.6%) / fr079 val 14.5% (was 14.8%) / MIT 27.3% / corridor 0.3% (was 1.0%)
fr079_train_12007.5% (was 8.8%) /intel_train_15016.5% (was 18.0%)
- ✅ 所見 (honest): naive trim on map + all pyramid levels regressed long train to ~23%; match-only + finest-level fix recovers and improves.
mit_train_120slight regression (31.6% vs 29.1%). - 状態: 実装済
- ✅ Robot-frame rolling local map — voxel merge + radius prune; reference polar profile rebuilt from accumulated points
- ✅ harness 有効化 (
runPSM:use_local_map=true, radius 15 m, voxel 0.15 m) - ✅ Public benchmark (local map):
- Intel 15.3% (was 21.8%) / fr079 val 14.3% (was 13.9%) / MIT 28.5% / corridor 4.4% (was 11.6%)
fr079_train_120046.7% (was 72.2%) /intel_train_15022.5% (was 23.7%)
- ✅ 所見 (honest): local map fixes Intel weakness; fr079 val ~flat; long train improves but still behind NDT/PL-ICP.
fr079_train_200honest negative (~681%). - 未実装: occupancy/co-occurrence matrix
- 状態: 実装済
- ✅ Robot-frame rolling local map — voxel merge + radius prune; CP spatial index; RR bearing-window (not beam index)
- ✅ harness 有効化 (
runIDC:use_local_map=true, radius 15 m, voxel 0.15 m) - ✅ Public benchmark (local map):
- Intel 15.2% (was 15.3%) / fr079 val 14.3% (was 27.7%) / MIT 27.8% / corridor 3.6% (was 42.6%)
fr079_train_120035.0% (was 46.4%) /intel_train_15025.1% (was 41.9%)
- ✅ 所見 (honest): local map fixes IDC's main weakness (fr079 + corridor). Intel ~flat. Long train still behind NDT/PL-ICP.
fr079_train_200honest negative (~85%). - 未実装: visibility filter、trimmed IDC
- 状態: 実装済
- ✅ 3-level multi-resolution pyramid — coarse→fine Gauss-Newton with
pyramid_scale=1.5 - ✅ harness 有効化 (
runNDT2D:pyramid_levels=3,pyramid_scale=1.5) - ✅ Public benchmark (pyramid + local map):
- Intel 14.6% / fr079 val 14.8% / MIT 28.1% / corridor 1.0%
fr079_train_12008.8% (was 10.3% P13) /intel_train_15018.0% (was 20.5%) /mit_train_12029.1% (best)
- ✅ 所見 (honest): mild pyramid (scale 1.5, 3 levels) recovers long-window drift on
fr079_train_1200. P17 outlier trim further improves to 7.5%. Naive trim regressed long train — see honest note above. - 未実装: —
- 状態: ✅ 実装済 (P17 outlier trim)
- ✅ Robot-frame rolling local map — voxel merge + radius prune (PL-ICP/MbICP パターン)
- ✅ harness 有効化 (
runNDT2D) - ✅ Public benchmark (local map):
- Intel 14.9% (was 14.8%) / fr079 14.4% (was 21.8%) / MIT 27.8% (was 29.2%) / corridor 0.8% (was 22.3%)
- ✅ 所見 (honest): local map は fr079 val + corridor で大幅改善。Intel は横ばい。
fr079_train_1200は scan-to-scan 7.4% → 10.3% (微悪化)。 - 未実装: outlier trimming
- 状態: 実装済
- ✅ 実装:
papers/ndt_2d/— grid Gaussian cells + Gauss-Newton NDT score (no correspondences)。 - ✅ 統合:
scan_dogfooding --methods ndt_2d - ✅ テスト:
test_ndt_2d5 cases PASS（pure translation で 3×3 cell + bilinear lookup 修正）。 - ✅ Public benchmark:
- Intel 14.8% (RF2O 14.3% に僅差) / fr079 21.8% / MIT 29.2% / corridor 22.3%
- ✅ 所見 (honest): real logs で RF2O 近傍 (correspondence-free) だが corridor 22.3% — PL-ICP 0.4% に大差。scan-to-scan single-resolution port。
- 未実装: multi-resolution pyramid、outlier trimming
- 状態: ✅ push 済 (
8ea40f1)
- ✅ 実装:
papers/idc/— CP (closest point) translation + RR (matching range) rotation fusion。 - ✅ 統合:
scan_dogfooding --methods idc - ✅ テスト:
test_idc7 cases PASS（pure translation threshold 0.2→0.25）。 - ✅ Public benchmark (local map, 2026-06-10):
- Intel 15.2% / fr079 14.3% / MIT 27.8% / corridor 3.6%
- ✅ 所見 (honest): local map で fr079/corridor 大幅改善。Intel mid-pack 維持。long train は NDT/PL-ICP 劣後。 visibility filter / trimmed IDC 無し。
- 未実装: trimmed IDC、visibility filter
- 状態: ✅ 実装済 (P15 local map)
- ✅ 実装:
papers/mb_icp/— Minguez/Lamiraux/Montesano の config-space metric ICP。 対応探索と Gauss-Newton 最小化の両方で||p||^2 + L^2による rotation-aware metric を使用。 - ✅ 統合:
scan_dogfooding --methods mb_icp、run_scan2d_benchmark.sh、CI smoke。 - ✅ テスト:
test_mb_icp5 cases PASS。 - ✅ Public benchmark (canonical JSON 8-method refresh):
- Intel 17.1% / fr079 16.6% / MIT 27.3% / corridor 0.46%
- ✅ 所見 (honest): fixture winner ではないが、fr079/MIT/corridor のバランスが良い。 PL-ICP より real logs で安定し、corridor では PL-ICP に近い。scan-to-scan のため長距離 drift は残る。 visibility rejection / local map / exact point-to-segment metric は未実装。
今回追加した主な差分:
| 領域 | 変更 |
|---|---|
| 実装 | papers/mb_icp/include/mb_icp/mb_icp.h, papers/mb_icp/src/mb_icp.cpp |
| テスト | papers/mb_icp/test/test_mb_icp.cpp — 初期化、並進、回転、並進+回転、pose2D |
| ビルド | papers/mb_icp/CMakeLists.txt, root CMakeLists.txt, evaluation/CMakeLists.txt |
| Harness | evaluation/src/scan_dogfooding.cpp — loader, runMbICP, selector mb_icp, JSON note |
| Batch/CI | evaluation/scripts/run_scan2d_benchmark.sh, evaluation/scripts/smoke_scan2d_fixture.sh |
| Docs/catalog | papers/mb_icp/README.md, README.md, docs/benchmarks/scan2d/README.md, docs/methods.json, public_bundle.json, SETUP_2D_SCAN_BENCHMARK.md |
| Artifacts | docs/benchmarks/scan2d/{intel_val_73,fr079_val_384,mit_val_33,rf2o_corridor}.json を 8-method canonical として再生成 |
実装メモ:
- reference scan は beam 隣接点から segment list を作る。range gap が
max_neighbor_gapを超える segment は切る。 - 対応探索は transformed current point を reference segment に射影し、
metricDistanceSquared()で最小の対応を選ぶ。 - metric は
delta^T Q delta相当で、Q = I - aa^T/(||ref||^2 + L^2)、a=(ref_y,-ref_x)。 これにより遠距離で回転に吸収できる tangential error を弱める。 - solve は同じ metric matrix を Gauss-Newton normal equation に入れる。対応集合は
trim_fraction=0.9で上位 90% を使用。 scan_dogfoodingの MbICP 既定はmax_metric_distance=1.5,metric_radius=1.0,max_neighbor_gap=2.0,trim_fraction=0.9。
パラメータ探索メモ:
metric_radius |
Intel | fr079 | MIT | Corridor | 所見 |
|---|---|---|---|---|---|
| 1.0 | 17.07 | 16.55 | 27.33 | 0.46 | 採用。4 fixture 平均 drift が最良 |
| 1.5 | 16.75 | 20.92 | 27.10 | 0.45 | fr079 が悪化 |
| 2.0 | 14.98 | 20.32 | 27.02 | 0.43 | Intel/MIT は良いが fr079 が悪化 |
| 3.0 | 15.83 | 20.18 | 27.83 | 0.41 | corridor は最良だが real logs で弱い |
採用値は single-fixture 最適ではなく、公開4 fixture の平均と honest leaderboard を優先。
次に retune する場合は public_bundle.json と docs/benchmarks/scan2d/README.md の数値も同時に更新する。
未実装 / 論文との差分:
- exact な MbICP point-to-segment metric derivation ではなく、segment 射影点に metric を評価する近似。
- visibility rejection、range discontinuity に基づく occlusion filter、local submap は未実装。
- scan-to-scan のみ。長い Bonn logs で drift が残るのは実装バグではなく current harness の設計限界。
検証済みコマンド:
rtk cmake -B build -DCMAKE_BUILD_TYPE=Release
rtk cmake --build build --target test_mb_icp scan_dogfooding -j$(nproc)
rtk ctest --test-dir build -R test_mb_icp --output-on-failure
rtk bash evaluation/scripts/run_scan2d_benchmark.sh
rtk jq empty docs/methods.json docs/benchmarks/scan2d/public_bundle.json \
docs/benchmarks/scan2d/intel_val_73.json \
docs/benchmarks/scan2d/fr079_val_384.json \
docs/benchmarks/scan2d/mit_val_33.json \
docs/benchmarks/scan2d/rf2o_corridor.json最終 commit 前にもう一度回す推奨:
rtk ctest --test-dir build -R 'test_mb_icp|test_rf2o|test_pl_icp|test_csm|test_kinematic_icp|test_psm|test_ndt_2d|test_idc' --output-on-failure
rtk bash evaluation/scripts/smoke_scan2d_fixture.sh
rtk python3 evaluation/scripts/validate_showcase.py --root .- ✅
scan_dogfoodingdrift 列: KITTI-style RPE (Drift [%])、--summary-json出力。 - ✅
prepare_2d_scan_inputs.py: ROS1sensor_msgs/LaserScan→ scan tree + GT 補間。 - ✅
prepare_bonn_2dslam_inputs.py: Bonn JSON → committed fixtures。 - ✅
SETUP_2D_SCAN_BENCHMARK.md: 合成 / bag / Bonn 手順 + 8 法 run 例。 - ✅
run_scan2d_benchmark.sh: 4 fixtures × 8 methods 一括 refresh。 - ✅
smoke_scan2d_fixture.sh: CI — Intel 20f × 8 methods (.github/workflows/ci.yml)。 - ✅
plot_scan2d_gt_overview.py: public GT 軌跡 PNG。 - ✅ Bonn fixtures (committed, GT = dataset odometry proxy):
| Fixture | Source | Frames | Beams | Traj [m] | Size |
|---|---|---|---|---|---|
intel_val_73 |
Bonn intel/val.json |
73 | 180 | ~378 | ~600 KB |
fr079_val_384 |
Bonn fr079/val.json |
384 | 360 | ~373 | ~3.1 MB |
mit_val_33 |
Bonn mit/val.json |
33 | 360 | ~267 | ~280 KB |
- ✅ Synthetic fixtures:
| Fixture | Frames | Traj [m] | 用途 |
|---|---|---|---|
rf2o_smoke |
60 | ~18 | 高速混合運動 smoke |
rf2o_corridor |
120 | ~9.5 | 低速 mixed motion (0.08 m/f) |
注意: 合成 corridor は bounded box raycast。純 parallel-wall corridor は 2D odom 退化 (forward structure 不足)。real hallway は ROS bag export を推奨。
- ✅ リーダーボード hub:
docs/benchmarks/scan2d/README.md
GT-seed on frame 0。Drift [%] — lower is better。
| # | Method | Intel val | fr079 val | MIT val | Synth corridor | Best on |
|---|---|---|---|---|---|---|
| 73f / 378m | 384f / 373m | 33f / 267m | 120f / 9.5m | |||
| 43 | RF2O | 14.3 | 15.4 | 27.6 | 1.3 | Intel |
| 48 | NDT-2D | 14.8 | 21.8 | 29.2 | 22.3 | — |
| 49 | IDC | 15.3 | 27.7 | 29.5 | 42.6 | — |
| 45 | CSM | 16.0 | 20.6 | 29.2 | 73.3 | — |
| 44 | PL-ICP | 15.0 | 14.1 | 27.2 | 0.01 | Corridor + fr079 competitive |
| 50 | MbICP | 17.1 | 16.6 | 27.3 | 0.5 | — |
| 46 | Kinematic-ICP | 18.4 | 18.9 | 23.4 | 83.8 | MIT (indicative) |
| 47 | PSM | 21.8 | 13.9 | 27.9 | 11.6 | fr079 |
- 単一 winner なし — fixture 依存が強い (PSM@fr079 vs RF2O@Intel vs PL-ICP@corridor)。
- scan-to-scan 限界 — 長 public log では range-flow (RF2O) か polar (PSM) が相対的に強いが どちらも GT proxy 対 drift 15–30% 帯。local map 無しでは Karto 論文数値には未達。
- 合成 vs real — corridor で PL-ICP 0.4% でも fr079 で 41% — 合成勝利は real 一般化しない。
- CSM DT — fr079 38.9%→20.6% は positive engineering だが corridor ~73% は honest negative 継続。
- Kinematic-ICP —
--wheel-odom-from-gt必須。real wheel odom 無し環境では skip/劣化。 - MIT val 33f — 全 drift indicative。ranking 参考程度。
| Fixture | Canonical JSON | 備考 |
|---|---|---|
| Intel | docs/benchmarks/scan2d/intel_val_73.json |
8-method canonical |
| fr079 | docs/benchmarks/scan2d/fr079_val_384.json |
8-method canonical |
| MIT | docs/benchmarks/scan2d/mit_val_33.json |
short window; indicative |
| Corridor | docs/benchmarks/scan2d/rf2o_corridor.json |
8-method canonical |
| Summary | docs/benchmarks/scan2d/public_bundle.json |
schema v2, 8 法 |
| Partial | *_idc.json, *_ndt2d.json, *_csm_dt.json |
method-specific reruns |
再現:
cmake --build build --target scan_dogfooding
bash evaluation/scripts/run_scan2d_benchmark.sh
python3 evaluation/scripts/plot_scan2d_gt_overview.py
bash evaluation/scripts/smoke_scan2d_fixture.sh # CI 同等ユーザ指示「markdown を整理」に基づき、2D scan 関連 docs を 単一ハブ に集約。
| ファイル | 変更 |
|---|---|
新規 docs/benchmarks/scan2d/README.md |
8×4 リーダーボード、artifact index、再現手順、所見 |
README.md |
Eight、MbICP 行、詳細はハブへ |
public_bundle.json |
schema v2、8 法 + corridor、drift 同期 |
SETUP_2D_SCAN_BENCHMARK.md |
8 法 run 例、CI smoke、ハブリンク |
papers/{rf2o,pl_icp,csm,psm,ndt_2d,idc,kinematic_icp,mb_icp}/README.md |
ベンチ表 or ハブリンク統一 |
PLAN.md |
本更新 |
ドキュメント階層:
README.md (1-screen 概要)
└── docs/benchmarks/scan2d/README.md (canonical hub)
├── public_bundle.json
├── intel_val_73.json / fr079_val_384.json / ...
└── papers/*/README.md (per-method honest notes)
未完了 (docs): なし — 2026-06-11 に methods.json 2D tag 強化 (local-map/distance-transform tag + opt-in 注記) と
validate_showcase.py scan2d セクション (public_bundle ↔ fixture JSON の drift 整合検証) を実施済。
| Priority | タスク | 状態 (2026-06-10) |
|---|---|---|
| P0 | MbICP + 8-method refresh 差分を最終検証して commit | ✅ 検証済 (8/8 tests, smoke, validate_showcase, benchmark refresh) |
| P1 | git push — IDC + CSM-DT + markdown + MbICP refresh |
ユーザ明示指示待ち |
| P2 | PL-ICP/MbICP local map 拡張 | ✅ MbICP robot-frame cache (Intel 14.5%, fr079 15.4%; ~2.3min) + PL-ICP robot-frame cache (Intel 15.0%, fr079 14.1%; ~26s) + stamp-indexed search speedup |
| P3 | Karto-style map matcher / spatial index for local map | ✅ karto_matcher + Olson coarse BnB; fr079 14.8% drift (Intel 15.1%); corridor honest negative |
| P4 | 長め MIT/Bonn window 追加 | ✅ mit_train_120 + intel_train_150 fixtures + run_scan2d_long_benchmark.sh |
| P5 | CSM robot-frame local map | ✅ Intel 14.5%, fr079 14.5%; corridor 95.8% (honest negative vs scan-to-scan 73%) |
| P6 | CI long-fixture smoke | ✅ smoke_scan2d_long_fixture.sh — MIT/Intel train 20f × 9 methods (.github/workflows/ci.yml) |
| P7 | CSM Olson coarse branch-and-bound | ✅ Karto から移植; fr079 14.9%, Intel 15.2%; corridor ~102% (honest negative) |
| P8 | fr079_train long window | ✅ fr079_train_200 fixture + benchmark + CI long smoke |
| P9 | CSM 速度チューニング | ✅ 64-node BnB + finest-only refine + score lookup; Intel 14.7% (~79 FPS), fr079 14.3% (~58 FPS), corridor 41% (was 102%) |
| P10 | Karto-Matcher への CSM 同チューニング移植 | ✅ 64-node BnB + finest-only refine + score lookup; Intel 14.7% (~64 FPS), fr079 14.3% (~47 FPS), corridor 41% (was 102%) |
| P11 | Felzenszwalb EDT (CSM + Karto) | ✅ common/felzenszwalb_edt; Intel 14.0%, fr079 13.7%, corridor 30.5% (was 41%); fr079_train_200 regressed (indicative) |
| P12 | fr079_train 更長 window | ✅ fr079_train_1200 (~150 m); CSM/Karto 17.6% (vs 40% on 200f); CI long smoke uses 1200 fixture |
| P13 | NDT-2D robot-frame local map | ✅ voxel merge + radius prune; fr079 14.4% (was 21.8%), corridor 0.8% (was 22.3%); Intel 14.9% |
| P14 | NDT-2D multi-resolution pyramid | ✅ 3-level coarse→fine (scale 1.5); fr079_train_1200 8.8% (was 10.3%); mit_train_120 29.1% (best) |
| P15 | IDC robot-frame local map | ✅ voxel merge + radius prune + bearing-window RR; fr079 val 14.3% (was 27.7%); corridor 3.6% (was 42.6%); Intel 15.2% |
| P16 | PSM robot-frame local map | ✅ point cache → polar profile rebuild; Intel 15.3% (was 21.8%); fr079 14.3%; corridor 4.4% (was 11.6%) |
| P17 | NDT-2D outlier trimming | ✅ match-only range jump + finest-level Mahalanobis trim; fr079_train_1200 7.5% (was 8.8%); corridor 0.3%; Intel ~flat |
| P18 | RF2O robot-frame local map | ✅ opt-in (default off) — val は fr079 13.9%/corridor 0.6% と改善するが long train が全 config 退行 (fr079_train_1200 10.6→20.1%+, fr079_train_200 30→88%+); min-range polar 投影が原因の family-level honest negative |
| P19 | Kinematic-ICP point 型 local map | ✅ opt-in (default off) — mit_train_120 46→13% / fr079_train_200 11→6% と効く窓がある一方、MIT val 23→31% (唯一のリーダー fixture) と fr079_train_1200 11→19% を破壊、safe shared config 無し |
| — | PG-LIO (3D) 改善 | 保留 (honest negative) |
| — | KITTI Odom full rerun | データ入手 |
Do NOT (明示指示なし):
- 50本目を未着手扱いに戻す
- PG-LIO を 2D より優先
- force push / git config 変更
- canonical JSON を partial artifact (
*_idc.json,*_csm_dt.json) の古い数字で上書き
§00.6c 以前の旧 2D 節 (00.53 RF2O 等) は上記 §00.6c-43〜50 に統合済み。以下は 3D 37 本目 VLOM の §00.53。
This section is the authoritative current handoff for June 2026 2D LiDAR campaign (§00.6c–§00.60). Older sections below still matter for 3D benchmark history, recipe provenance, and paper-grade claims, but they describe the May 2026 / early-June 3D research state. The active direction is:
- 2D scan odometry — 8-method leaderboard is current; next value is local mapping / Karto-style matching.
- OSS polish — showcase/demo/CI remain important (§0, validate_showcase contract).
- 3D on hold — PG-LIO honest negative; KITTI full reruns blocked by data.
The user's current theme has been:
- "shibaraku 2D LiDAR de yatte iku" (2D campaign pivot)
- "omosiroi kihatu wo siyou. star wo huyasitai." (interesting mechanisms + stars)
- "ittan plan md wo naganaga kousin" (long PLAN.md refresh — this update)
Interpretation: this repo should look and feel like a strong OSS project, not only a private research scratchpad. First-clone users should immediately see a usable method explorer, a runnable demo, a generated report, and CI-backed proof that many methods can be validated together.
This subsection supersedes 0.1–0.7 where they conflict; those describe the pre-merge dirty-worktree state.
- Showcase merged to
main. PR #2 (wip/profile-expansion-refresh→main) was merged (merge commit4ddb7c3). The merge brought inmain's 14 commits (HDL-400 B benchmark docs,run_local_evaluation_suite.sh/eval_local_suiteCMake target, Ceres/OpenCV build tweaks, KISS-ICP empty-frame + common golden tests). Conflicts were resolved keeping the branch's newer research/showcase state for data/docs and merging both sides for source (run_experiment_matrix.pykept the lower-caseis_metric_validtaxonomy + new--merge-existing-index; the duplicateproblem_run_from_aggregatewas disambiguated intoproblem_run_from_aggregate(manifest form) andproblem_run_from_aggregate_file(aggregate-only form)). - Method count 33 → 35. Two new from-scratch LiDAR ports were added:
papers/genz_icp/— GenZ-ICP style degeneracy-robust odometry (adaptive point-to-plane / point-to-point hybrid, normal+planarity estimation on the KISS-style voxel map). Selectorgenz_icp, profiles--genz-fast-profile/--genz-dense-profile/--genz-planarity-threshold.test_genz_icp.papers/rko_lio/— RKO-LIO style scan-to-map odometry with an IMU gyro rotation prior (viaimu_preintegration), constant-velocity fallback when noimu.csv. Selectorrko_lio.test_rko_lio. Always runs (does not skip); note text records whether the IMU prior was used.- Both have
docs/methods.jsonentries (catalog now 41 entries = 41papers/*).
- Drift column.
pcd_dogfoodingresults table now printsDrift[m/100m](the existing KITTI-stylerpe_trans_pct, previously only in summary JSON; shows-for trajectories shorter than the RPE segment). - GenZ-ICP KITTI Odom 108-window finding (memory
genz_icp_vs_kiss_kitti_windows.md): GenZ-ICP beats KISS-ICP on drift in 3/5 windows (00/02/07) but is not universal (loses on 05), is slower, and does not rescue the seq 08 no-seed divergence (~33 m/100m, like KISS). NDT (GT-seeded) remains the overall window winner. RKO-LIO full benchmark intentionally not yet run (user deferred). - Commits on branch after the merge:
dd69400(GenZ-ICP),e1ecc6a(drift column),3645d67(RKO-LIO). Branch pushed to origin. - Verification:
cmake --build build --target pcd_dogfooding multimodal_dogfoodingclean;ctest -R "genz_icp|rko_lio"pass;python3 -m unittest discover -s tests→ 49 pass; one-command demo (broad) +validate_showcase.py --require-demovalid. The demo all-OK profiles were intentionally not changed (genz_icp/rko_lio are available via--methodsbut not in the default broad/full sets). - NCLT dataset ingested (no-form, S3 direct).
evaluation/scripts/prepare_nclt_inputs.pySETUP_NCLT_BENCHMARK.mdconvert NCLT (UMich, Velodyne+MS25 IMU+6-DoF GT) into the standarddogfooding_results/nclt_*(cloud.pcd + frame_timestamps.csv + imu.csv) andexperiments/reference_data/nclt_*_gt.csv. Smallest session2013-01-10(velodyne 2.9 GB) ingested at 120- and 600-frame windows. Raw download lives outside the repo innclt_raw/; the PCD trees are gitignored, only GT CSVs are committed.
- Active manifests
rko_lio_nclt_2013_01_10(gyro-bias-gain sweep) andgenz_icp_nclt_2013_01_10, run throughrun_experiment_matrix.py --merge-existing-indexand merged intoindex.json+ generated docs. - Finding (memory
nclt_dataset_ingested.md): on NCLT the raw IMU rotation prior (gyro_bias_gain 0) is best (RKO-LIO ATE 0.141 @120, 0.385 @600 — beating FAST-LIO2's 0.469 @600, 2nd only to GT-seeded NDT). Optimal bias gain is dataset-dependent — HDL-400 wants 0.3 (high IMU bias), NCLT wants 0 (low-bias MS25); no universal default. KISS-ICP (no IMU) drifts to 7.2 m @600. NCLT is where the inertial prior clearly pays off. - Bug fixed:
variant_result_from_dictinrun_experiment_matrix.py(main-side helper pulled in by the merge) omitted the branch's requiredrpe_trans_pct/rpe_rot_deg_per_mfields, breaking the--merge-existing-indexpath. Now fixed.
0.0b Update — NCLT cross-method + full-trajectory benchmark + LiTAMIN2 improvement (2026-06-02, later session)
-
Full NCLT tree generated.
prepare_nclt_inputs.py --max-frames -1produceddogfooding_results/nclt_2013_01_10_full(5105 GT-matched frames) + committedexperiments/reference_data/nclt_2013_01_10_full_gt.csv. Passingmax_framestopcd_dogfoodinglets any prefix window be evaluated off the full tree (2000-frame windows are the practical tuning size — they reproduce the long-trajectory drift while staying fast). -
Mechanism clarified. In this dogfooding tool, GT-seeded methods (NDT, LiTAMIN2, GICP-family) use a per-frame seed:
T_init_guess = applySeedPerturbation(gt[i], …), then accept the refinement only if it stays within the gate (NDT 1.5 m/0.2 rad, others 2.0 m/0.25 rad), else roll back togt[i](weak-update fallback). So their ATE is the per-frame registration residual from GT, not accumulated odometry drift. RKO-LIO / FAST-LIO2 / KISS-ICP only seed the first pose → they are odometry and drift over the full trajectory (a different category from the per-frame-seeded registration methods). -
600-frame horizontal comparison (11 methods). Winner NDT 0.198, then LiTAMIN2 0.380 ≈ RKO-LIO(gain0) 0.385, FAST-LIO2 0.469, small_gicp 1.024; no-seed odometry blows up (KISS 7.25, CT-ICP 13.6, GenZ 26.0, Point-LIO 122).
-
Full 5105-frame benchmark — scale-dependent reversal (key finding).
method (full 5105) seed ATE [m] drift [m/100m] NDT GT 0.122 0.460 LiTAMIN2 voxel0.5/iter12 GT 0.582 1.315 small_gicp voxel0.5 GT 1.040 2.786 small_gicp default GT 1.086 2.288 LiTAMIN2 default (voxel2.0) GT 1.149 2.492 RKO-LIO (gain0) init 7.334 2.077 RKO-LIO (gain0.3) init 24.43 4.568 FAST-LIO2 – 16.126 3.330 KISS-ICP – 60.629 15.134 NDT actually improves at full scale (0.198 → 0.122) thanks to its tighter gate + GT reanchoring. The 600-frame near-ties (LiTAMIN2/RKO-LIO ≈ NDT) do not hold at full.
-
Improvement delivered: LiTAMIN2 voxel 2.0 → 0.5 (+ iter 12). NCLT's HDL-32E scans are sparser than KITTI's HDL-64E, so the default voxel 2.0 is too coarse. Transferring the KITTI cluster-T1 recipe (voxel 0.5 + iter 12) cuts full ATE 1.149 → 0.582 (−49 %) and drift 2.492 → 1.315 (−47 %), moving LiTAMIN2 to a clear #2 behind NDT. voxel 1.0 does not help (1.044 @2000) — fineness is the lever. Captured as the reproducible manifest
experiments/litamin2_nclt_2013_01_10_matrix.json(voxel sweep, merged intoindex.json). -
Negative results recorded honestly: small_gicp does not benefit from the voxel-0.5 lever (full 1.086 → 1.040 ATE but drift worsens 2.288 → 2.786); RKO-LIO gyro-bias gain 0.3 is worse than gain 0 at full (24.4 vs 7.3) — bias correction over-integrates on the long trajectory. Optimal gain is both dataset- and scale-dependent; no universal default.
-
Second NCLT session ingested for cross-session validation. Downloaded 2012-12-01 (velodyne 13.4 GB, fetched via 8-way HTTP range requests at ~12 MB/s after the single stream stalled at ~1 MB/s), ingested a 5000-frame / 937 m prefix as
dogfooding_results/nclt_2012_12_01_5000+ committed…_5000_gt.csv. -
LiTAMIN2 T1 (voxel0.5/iter12) generalizes across NCLT sessions. On 2012-12-01 it cuts ATE 0.869 → 0.519 (−40 %) and drift 1.844 → 1.357 (−26 %), mirroring the 2013-01-10 full-5105 −49 %/−47 %. NDT is the winner again (0.118, vs 2013-01-10's 0.122). The fine-voxel recipe is a session-independent NCLT improvement, not a per-session artifact. Manifest
experiments/litamin2_nclt_2012_12_01_matrix.json(merged intoindex.json).method (2012-12-01, 5000 f) ATE [m] drift [m/100m] NDT 0.118 0.374 LiTAMIN2 T1 (voxel0.5/iter12) 0.519 1.357 LiTAMIN2 default (voxel2.0) 0.869 1.844 small_gicp default 1.049 2.616 small_gicp cd1.5 1.009 2.883 -
small_gicp #3 improvement: the seed gate is the lever, not voxel/correspondence. The CLI levers that fix LiTAMIN2 wash out for small_gicp at full scale: correspondence distance 1.5 gives a real −23 % at 2000 frames (1.181 → 0.908) but evaporates by full-5105 (1.086 → 1.089); voxel0.5 is likewise flat. The actual bottleneck is the weak-update seed gate, which defaulted to 2.0 m / 0.25 rad (looser than NDT's 1.5 m / 0.2 rad) and was not CLI-exposed. Added
--small-gicp-max-seed-translation-delta/-rotation-delta-rad(mirrors the existing CT-ICP flags) + a seed-fallback counter inrunSmallGICP. Tightening the gate monotonically cuts full-5105 ATE:gate (full 5105) ATE [m] drift GT-fallback % (2000 f) 2.0 m / 0.25 rad (default) 1.086 2.288 2.5 % 1.5 m / 0.2 rad (NDT-equal) 0.882 2.357 4.2 % 1.0 m / 0.15 rad (balanced) 0.675 2.036 9.3 % 0.5 m / 0.1 rad (aggressive) 0.348 1.313 20.2 % At 1.0 m / 0.15 rad ATE drops −38 % with only 9.3 % GT fallback (the recommended balanced config); at 0.5 m / 0.1 rad it drops −68 % and overtakes LiTAMIN2 T1 (0.582) for NCLT #2. Honest caveat: a tighter gate also raises the GT-fallback rate (2.5 % → 20.2 %), so part of the gain is more aggressive reversion to the exact GT seed on the worst registrations, not purely better refinement — the same weak-update mechanism NDT benefits from, now exposed for small_gicp. The fair fixed-gate comparison (small_gicp 0.882 vs NDT 0.122 at the identical 1.5 m gate) shows NDT's edge is genuine refinement quality, not just the gate. Manifest
experiments/small_gicp_nclt_2013_01_10_matrix.json.
| Item | Value |
|---|---|
| Branch | wip/profile-expansion-refresh |
| HEAD at update time | 947912d |
| Worktree | Dirty by design; current showcase/demo changes are not committed yet |
| Local instruction | Prefix shell commands with rtk when operating as Codex |
| Main current artifact | One-command demo + GitHub Pages showcase + validators |
| Current generated demo output | Ignored under experiments/results/runs/demo_localization_zoo/ |
Current dirty/untracked files at the time of this PLAN update:
- Modified:
.github/workflows/ci.ymlREADME.mddocs/assets/site.cssdocs/index.htmltests/test_experiment_scripts.py
- Untracked:
docs/assets/explorer_preview.pngdocs/methods.jsonevaluation/scripts/demo_localization_zoo.shevaluation/scripts/generate_demo_report.pyevaluation/scripts/validate_demo_artifacts.pyevaluation/scripts/validate_showcase.py
The generated demo run artifacts are intentionally not tracked. They are produced by the demo script and uploaded by CI as an artifact:
experiments/results/runs/demo_localization_zoo/report.htmlexperiments/results/runs/demo_localization_zoo/manifest.jsonexperiments/results/runs/demo_localization_zoo/synthetic_benchmark.logexperiments/results/runs/demo_localization_zoo/lidar_fixture_summary.jsonexperiments/results/runs/demo_localization_zoo/multimodal_fixture_summary.json- trajectory text outputs under
benchmark_results/anddogfooding_results/
The repo now has a public-facing showcase layer in addition to the older research benchmark layer.
docs/index.html has been turned into a richer interactive
front page:
- Loads latest benchmark data from
docs/benchmarks/latest/results.json. - Loads method metadata from the new
docs/methods.jsoninstead of hardcoded JavaScript arrays. - Renders a method explorer covering every
papers/*directory. - Shows starter tracks for different user intents.
- Shows a benchmark scatter/leaderboard for the committed latest snapshot.
- Includes OpenGraph/Twitter metadata for link previews.
- Points users toward the one-command local demo.
docs/assets/site.css was redesigned for the explorer:
- More dashboard-like and scan-friendly.
- Responsive layout for desktop/mobile.
- No giant marketing-only hero; the page is meant to be useful immediately.
docs/assets/explorer_preview.png is a new
preview image referenced from the README and validated as a PNG by the showcase
validator.
docs/methods.json is new:
schema_version: 1- 39 method entries, one for each
papers/*directory. - Required fields per method:
namefamilyscopesignalshrefsummarytags
- 4 starter tracks:
- quick start / first run
- accuracy oriented
- fusion / multimodal
- degeneracy / robustness
The method catalog is now test-covered. If a new paper directory is added without a catalog entry, the Python tests fail.
evaluation/scripts/demo_localization_zoo.sh
is new and is now the first-run path for users. It:
- Builds the C++ targets unless
--skip-buildis provided. - Runs
synthetic_benchmark. - Runs the committed three-frame MCD fixture through selected LiDAR methods.
- Runs the same fixture through selected multimodal methods.
- Writes logs, summary JSON, trajectories,
report.html, andmanifest.json. - Calls the demo artifact validator at the end, so the command proves its own output contract.
The demo now has method profiles:
| Profile | LiDAR methods | Multimodal methods | Intended use |
|---|---|---|---|
quick |
4 | 2 | Fast old-style local loop |
broad |
24 | 6 | Default; best first-clone OSS proof |
full |
25 | 6 | Adds LiDAR FAST-LIO2 fixture validation |
Current LiDAR quick:
litamin2,gicp,ndt,kiss_icp
Current LiDAR broad:
litamin2,gicp,small_gicp,voxel_gicp,ndt,kiss_icp,dlo,dlio,aloam,floam,lego_loam,mulls,ct_icp,xicp,hdl_graph_slam,vgicp_slam,suma,balm2,isc_loam,loam_livox,lio_sam,lins,fast_lio_slam,point_lio
Current LiDAR full:
litamin2,gicp,small_gicp,voxel_gicp,ndt,kiss_icp,dlo,dlio,aloam,floam,lego_loam,mulls,ct_icp,xicp,hdl_graph_slam,vgicp_slam,suma,balm2,isc_loam,loam_livox,lio_sam,lins,fast_lio_slam,point_lio,fast_lio2
Note: full is not literally every LiDAR selector in pcd_dogfooding. It is
the largest all-OK set for the committed MCD fixture without requiring IMU-only
methods that skip when imu.csv is absent. ct_lio and clins currently skip
on this fixture because there is no synchronized IMU CSV. Do not include them
in the default all-OK profile unless the fixture gains IMU data or the validator
learns about expected skips.
Current multimodal quick:
okvis,fast_livo2
Current multimodal broad / full:
vins_fusion,okvis,orb_slam3,lvi_sam,fast_livo2,r2live
Useful invocations:
bash evaluation/scripts/demo_localization_zoo.sh
bash evaluation/scripts/demo_localization_zoo.sh --skip-build
bash evaluation/scripts/demo_localization_zoo.sh --skip-build --profile quick
bash evaluation/scripts/demo_localization_zoo.sh --skip-build --profile broad
bash evaluation/scripts/demo_localization_zoo.sh --skip-build --profile full
bash evaluation/scripts/demo_localization_zoo.sh --skip-build --methods litamin2,gicp --multimodal-methods okvisAs Codex in this workspace, remember to run those through rtk, e.g.
rtk bash evaluation/scripts/demo_localization_zoo.sh --skip-build.
evaluation/scripts/generate_demo_report.py
is new:
- Standard-library-only report generator.
- Parses synthetic benchmark logs.
- Reads LiDAR and multimodal summary JSON.
- Reads trajectory text files.
- Produces self-contained
report.htmlwith inline CSS/SVG. - Produces
manifest.jsonwith:- schema version
- command
- profile
- requested LiDAR method list
- requested multimodal method list
- actual method counts and statuses
- artifact paths
The manifest is not just a pretty output. It is now the contract between the demo script, validator, CI artifact upload, README claims, and showcase tests.
evaluation/scripts/validate_demo_artifacts.py
is new:
- Validates
manifest.json. - Validates
report.htmlrequired snippets. - Validates synthetic/LiDAR/multimodal summary artifacts.
- Requires every selected method to have
status == ok. - Validates exact method-set match by normalized method name.
- Can read expected method lists from CLI or from the manifest.
- Supports
--skip-multimodal.
Important behavior: if someone silently shrinks the default profile from 24 LiDAR methods to 4, the validator catches it when expected methods are provided or when manifest-vs-summary diverges.
evaluation/scripts/validate_showcase.py
is new:
- Validates README links and required snippets.
- Validates the preview PNG signature.
- Validates
docs/index.htmlsnippets and repo link. - Validates
docs/methods.jsoncoverage againstpapers/*. - Validates starter-track references.
- Validates the latest benchmark snapshot and trajectory plot.
- Optionally validates generated demo artifacts with
--require-demo.
.github/workflows/ci.yml now includes:
- One-command demo report generation after existing smoke checks.
- Showcase validation with
--require-demo. - Upload of
experiments/results/runs/demo_localization_zoo/as thelocalization-zoo-demo-reportartifact.
The CI path intentionally demonstrates the same thing a new user sees locally: build, run a real committed fixture, generate a report, validate it, and keep the report downloadable.
The following commands passed during the 2026-06-02 update:
rtk bash -n evaluation/scripts/demo_localization_zoo.sh
rtk python3 -m py_compile evaluation/scripts/generate_demo_report.py evaluation/scripts/validate_demo_artifacts.py evaluation/scripts/validate_showcase.py
rtk python3 evaluation/scripts/validate_showcase.py --skip-demo
rtk bash evaluation/scripts/demo_localization_zoo.sh --skip-build
rtk bash evaluation/scripts/demo_localization_zoo.sh --skip-build --profile full
rtk python3 evaluation/scripts/validate_showcase.py --require-demo --demo-dir experiments/results/runs/demo_localization_zoo
rtk python3 -m unittest discover -s tests -p 'test_*.py' -v
rtk proxy git diff --check -- README.md docs/index.html evaluation/scripts/demo_localization_zoo.sh evaluation/scripts/generate_demo_report.py evaluation/scripts/validate_demo_artifacts.py evaluation/scripts/validate_showcase.py tests/test_experiment_scripts.pyObserved validation result:
broaddemo: LiDAR 24 / multimodal 6 all OK.fulldemo: LiDAR 25 / multimodal 6 all OK.- Showcase validation: OK.
- Python tests: 49 tests OK.
- Whitespace check: OK.
Because full was the last demo run before this PLAN update, the ignored local
experiments/results/runs/demo_localization_zoo/manifest.json currently records
profile: full, lidar_fixture.method_count: 25, and
multimodal_fixture.method_count: 6.
tests/test_experiment_scripts.py now
includes three new showcase/demo-focused groups:
MethodCatalogTestsdocs/methods.jsonschema is version 1.- Catalog has unique method names and hrefs.
- Catalog hrefs exactly cover all
papers/*directories. - Required fields are present.
- Each referenced method README exists.
- Starter tracks reference known methods.
ShowcaseContractTests- Runs
validate_showcase.py --skip-demo.
- Runs
DemoReportScriptTests- Builds a minimal fake demo directory.
- Runs
generate_demo_report.py. - Checks report snippets.
- Checks manifest profile and requested method lists.
- Runs
validate_demo_artifacts.py. - Verifies an expected-method mismatch fails.
The README and Pages story should now be:
- This repo contains many localization/SLAM method ports.
- Users can browse the catalog on GitHub Pages.
- Users can run a one-command demo after clone.
- The demo validates a committed real-data MCD fixture, not just synthetic data.
- The default path compares many methods together.
- CI runs that same path and uploads the HTML report.
This is intentionally a stronger OSS adoption story than "read a long paper doc and manually pick scripts."
Keep these distinctions sharp:
- The committed MCD fixture is a smoke/demo fixture. It proves integration and output contracts, not paper-grade accuracy.
- The
broadandfullprofiles are all-OK fixture profiles, not a statement that every method is production-ready. - The demo report is reproducible from the repo because the fixture is committed.
- Full dataset benchmark claims still come from
experiments/results/*.json,docs/reproduction_status.md, and the older benchmark matrix. - Do not claim exact original-paper reproduction unless the taxonomy/docs say so.
- Do not include methods that skip on the fixture in default all-OK profiles unless the validator is changed to support expected skips with explicit reasons.
Recommended order:
-
Review the new untracked files as if preparing a commit.
- Confirm no generated demo outputs are accidentally staged.
- Confirm
docs/assets/explorer_preview.pngis intentionally tracked. - Confirm
docs/methods.jsonnames/summaries read well enough for public Pages.
-
Run one final local verification before commit.
rtk bash evaluation/scripts/demo_localization_zoo.sh --skip-buildrtk python3 evaluation/scripts/validate_showcase.py --require-demo --demo-dir experiments/results/runs/demo_localization_zoortk python3 -m unittest discover -s tests -p 'test_*.py' -vrtk proxy git diff --check -- README.md docs/index.html evaluation/scripts/demo_localization_zoo.sh evaluation/scripts/generate_demo_report.py evaluation/scripts/validate_demo_artifacts.py evaluation/scripts/validate_showcase.py tests/test_experiment_scripts.py PLAN.md
-
Commit the showcase/demo expansion.
- Suggested commit scope: README, docs page/assets, method catalog, demo scripts, validators, CI, tests, PLAN.
- Suggested commit message:
Add showcase explorer and broad demo validation
-
Then consider one star-growth follow-up, not all at once.
- Add README badges for CI / Pages / license.
- Add a compact "What runs in 60 seconds?" section.
- Add a GIF or screenshot from
report.html. - Add issue templates for "method request" and "dataset request".
- Add a
CONTRIBUTING.mdfocused on adding a method topapers/*anddocs/methods.json.
-
If expanding validation again, prefer expected-skip semantics before adding IMU-only selectors.
- Current fixture lacks
imu.csv. ct_lioandclinsskip correctly on this data.- A robust next step would let the manifest encode expected skips with reason text, but this is a different contract from the current all-OK demo.
- Current fixture lacks
Avoid these unless the user explicitly asks:
- Do not start a new paper-writing pass.
- Do not rerun large KITTI/MulRan full-sequence sweeps just to improve the README.
- Do not add more untracked generated artifacts.
- Do not collapse the old benchmark/reproduction docs into the new showcase page; the showcase should point into them, not replace them.
- Do not loosen validators to make demos pass. If a method is not OK, either remove it from an all-OK profile or add an explicit expected-skip contract.
⚠️ 2026-06-09: §00 が最新 handoff。 本章 (§1–§14) は 2026-05〜06 初旬の 3D benchmark / manifest / reproduction 状態の歴史的記録。現アクティブ作業は 2D LiDAR scan odometry (papers 43–50, §00.6c–§00.60)。git/論文数/次タスクは §00.2 を正とする。
| Item | Value |
|---|---|
| Branch | wip/profile-expansion-refresh |
| HEAD | 458c81a |
| Worktree | clean after NDT + LOAM-family KITTI Odom checkpoint commit |
| Indexed manifests | 392 |
| Indexed ready | 378 |
| Indexed blocked | 1 |
| Indexed skipped | 13 |
| Pending manifests | 200 |
| LiDAR families | 27 |
| Camera-aware families | 6 |
| Total active selectors | 33 |
| Python tests | 14/14 pass (last full run; not rerun after NDT/LOAM-family docs/artifact refresh) |
NDT T1 transfer confirmation on KITTI Odom seq 02/05/08 full. --ndt-resolution 0.5 --ndt-max-iterations 12 stayed sub-10cm on all three held-out long sequences: seq 02 = 0.0585 m ATE, seq 05 = 0.0594 m ATE, seq 08 = 0.0761 m ATE. Together with the existing seq 00 r05+i12 result (0.0707 m) and seq 07 r05+i12 result (0.0763 m), NDT now has a seeded KITTI Odom 5/5 sub-8cm T1-family recipe. The direct confirmation runs were slow on this export path (0.23-0.25 FPS), so the result is an accuracy/universality confirmation rather than a throughput recommendation.
A-LOAM KITTI Odom full transfer check: added aloam_kitti_seq_07_full_sweep plus seq 00/02/05/08 transfer matrices. Seq 07 fast wins the combined benchmark by throughput (4.0307 m ATE, 0.691% RPE, 7.26 FPS), while kitti_default is the seq 07 accuracy winner (2.5052 m ATE, 0.605% RPE, 3.26 FPS). Transfer is non-universal: accuracy winner is kitti_default on all held-out seqs, with ATE 9.5206 m (00), 50.7898 m (02), 5.1927 m (05), 18.6614 m (08). fast is faster but less accurate on all held-out seqs: 19.3711 m (00), 74.9663 m (02), 7.7893 m (05), 23.3578 m (08). Treat A-LOAM as a drift-level baseline, not a sub-meter KITTI Odom recipe.
F-LOAM KITTI Odom seq 07 full cluster probe: added floam_kitti_seq_07_full_sweep. dense is the accuracy/RPE winner (3.1736 m ATE, 0.590% RPE, 3.74 FPS), kitti_default is close in accuracy but much faster (3.2455 m, 0.789% RPE, 13.23 FPS), and fast wins the combined benchmark by throughput (5.0260 m, 1.262% RPE, 29.84 FPS). F-LOAM is faster than A-LOAM on this sequence but still drift-level; transfer to seq 00/02/05/08 remains unverified.
LeGO-LOAM KITTI Odom seq 07 full cluster probe: added lego_loam_kitti_seq_07_full_sweep. kitti_default is the accuracy/RPE winner (2.5579 m ATE, 0.527% RPE, 3.89 FPS), fast wins the combined benchmark by throughput (4.3850 m, 0.597% RPE, 10.03 FPS), and dense underperformed both (3.9454 m, 0.650% RPE, 3.68 FPS). LeGO-LOAM is the strongest LOAM-family accuracy profile on seq 07 so far, slightly worse than A-LOAM kitti_default in ATE but better in RPE; held-out transfer remains unverified.
MULLS KITTI Odom seq 07 full cluster probe: added mulls_kitti_seq_07_full_sweep. dense is the accuracy/RPE winner (8.2878 m ATE, 2.640% RPE, 1.27 FPS), kitti_default is close but slower/worse (8.4591 m, 2.720% RPE, 1.42 FPS), and fast wins the combined benchmark by throughput (10.5014 m, 2.994% RPE, 4.13 FPS). MULLS is the weakest LOAM-family seq 07 candidate so far and very slow on full KITTI Odom; do not spend transfer budget on it unless specifically needed as a negative baseline.
F-LOAM + LeGO-LOAM held-out transfer: added seq 00/02/05/08 full transfer matrices for F-LOAM kitti_default/dense and LeGO-LOAM kitti_default. No universal LOAM-family recipe emerged. Per-seq LOAM accuracy winners are mixed: seq 00 = F-LOAM dense (8.5561 m ATE, 0.991% RPE, 3.37 FPS), seq 02 = LeGO-LOAM kitti_default (42.2088 m, 0.883% RPE, 3.79 FPS), seq 05 = A-LOAM kitti_default (5.1927 m, 0.512% RPE, 2.26 FPS; LeGO is nearly tied at 5.2182 m), seq 08 = F-LOAM kitti_default (16.6614 m, 1.566% RPE, 12.83 FPS). These improve over A-LOAM on seq 00/02/08 but remain drift-level and far from NDT/LiTAMIN2/GICP seeded winners.
GICP family recipe discovery + seed-dependence verification across all KITTI Odom sequences. State delta: 336 → 371 indexed manifests (35 new), HEAD d22a172 → 458c81a.
GICP family findings (memory entries: gicp_family_seq_07_recipe_divergence.md, small_gicp_fast_kitti_universal.md, small_gicp_seed_dependence.md):
- small_gicp
--small-gicp-fast-profileは KITTI Odom 5/5 + KITTI Raw 4/4 + MCD KTH で ATE winner = 9/10 scenes universal sub-meter (0.68-0.98 m). LiTAMIN2 T1 と並ぶ唯一の cross-KITTI universal recipe。indoor static (MCD NTU/TUHH) のみ dense_profile に転換。 - voxel_gicp
--voxel-gicp-dense-profileは KITTI Odom 5/5 で both ATE & RPE universal winner (ATE 0.94-1.05 m, RPE 1.5-1.8%). 強い but KITTI Raw / MCD では dataset-dependent (4/9 only). - KISS-ICP は long-trajectory で catastrophic: seq 02 で 39.23 m, seq 00 で 11.98 m, seq 08 で 19.41 m。dense_profile が 4/5 で best ATE recipe も全体的に他 method 比 10-30× worse on long-traj. Local-map-only architecture が drift-bound.
- LiTAMIN2 T1 (voxel=0.5+iter=12) transfer は GICP family で non-universal: KISS-ICP seq 07 のみ局所勝利、small_gicp で ATE は fast に負ける、voxel_gicp で全 seq 退行。T1 transfer は LiTAMIN2 専用 recipe.
Seed-dependence critical finding (memory entry: small_gicp_seed_dependence.md):
KITTI Odom seq 00 full no-seed test で 4 GICP family methods + LiTAMIN2 全てが完全発散:
| method | seeded ATE | no-seed ATE | drift |
|---|---|---|---|
| LiTAMIN2 T1 | 0.731 | ~110 m | +15,000% |
| small_gicp fast | 0.890 | 202.72 | +22,800% |
| voxel_gicp dense | 1.047 | 87.91 | +8,300% |
| CT-ICP best (cluster A) | 4.91 (seeded) | 12.69 | 唯一 functional |
→ production deployment (no-seed) on KITTI Odom long-trajectory では CT-ICP が唯一の選択肢。seeded benchmark での compositional ranking (LiTAMIN2 T1 / small_gicp fast / voxel_gicp dense / CT-ICP) が完全 reversed.
Earlier (2026-05-18..19, LiTAMIN2 saturation + CT-ICP completion):
LiTAMIN2 cluster T1 (voxel=0.5 + iter=12 + GT seed) は KITTI Odom 5/5 + MulRan 2/2 + MCD 3/3 + KITTI Raw 4/4 = 11/12 universal winner, 1/12 tied (noise floor 0.5 m). CT-ICP 比 1.8-70× dominance on seeded benchmark.
Earlier CT-ICP findings (memory entries: ct_icp_kitti_full_per_seq_best.md, ct_icp_cluster_a_cross_dataset_transfer.md, ct_icp_gt_seed_dataset_dependence.md):
- 5-cluster structure for CT-ICP recipes: cluster A (
map=50 + c2f σ×2) wins KITTI seq 00 (12.69 m) / 05 (7.76 m) / 08 (27.85 m). cluster B = A + corr=4 for seq 05. cluster C =bare + corr=8for seq 02 (50.63 m). cluster D =ms_chol + map=20for seq 07 (1.61 m) and KITTI Raw 0061 full (4.50 m). KITTI Raw 0009 is its own balanced-only family. - Cross-dataset transfer: cluster A + GT seed wins on MulRan parkinglot full (9.19 m, -36% vs prior best) and parkinglot 120 (2.55 m, -84%). cluster A + seed is the universal seeded winner on MCD KTH/TUHH/NTU/MulRan parkinglot — but not on KITTI Odom seq 07 where cluster D dominates seed-independently.
- GT seed dataset-dependence: seed helps drift-limited scenes (-44 to -88% on MulRan parkinglot, MCD KTH/TUHH) but hurts self-anchoring scenes (MCD NTU +39%, KITTI seq 07 neutral). On KITTI seq 00 full it produces the sharpest ATE/RPE flip yet observed: ATE -61% (12.69 → 4.91 m) but RPE +174% (2.10 → 5.76%).
- Knob axes exhaustively mapped on seq 00 cluster A: cauchy_coarse_mult (2.0 winner), cauchy_fine_sigma (default 0.5 winner), coarse_search_radius (2 winner, 4 ties), coarse_iter (2 winner), map_size (50 winner), corr_dist (default 100 winner). 2-D cauchy plane has its true minimum at (fine=0.5, coarse_mult=2.0).
Earlier (2026-05-17) foundational taxonomy layer (still in effect):
- Claim-level taxonomy:
reproduced > approximately_reproduced > indicative > smoke > ported. Bumped intoevaluation/data/paper_reported_numbers.json(schema v3) and rendered intodocs/reproduction_status.mdas a legend column. - Budget profile contract:
docs/budget_profiles.mddefinessmoke_200f_120s,practical_full_300s,extended_full_1800s,reference_full_unbounded. Manifests reference them viaproblem.budget_profile. - Family registry:
experiments/families.jsonclassifies the 33 method families intocore/extendedtiers andmaintained/timeout_prone/input_constrained/legacy/experimentalstatus. - Status taxonomy migration:
docs/status_taxonomy.mddefines the target per-variant enum. C++ binaries (pcd_dogfooding,multimodal_dogfooding) and the runner now emit the lower-case enum (ok,skipped,timeout_budget). Reserved values (tracking_lost,init_failed,input_unsupported,metric_invalid,no_gt) have field space inMethodResult::statusbut no detection logic yet. Legacy uppercase (OK/SKIPPED/TIMED_OUT) is normalized at ingest.
Earlier session work (already in main history):
- MulRan / Newer College benchmark scaffolding (commit
aee7611). - Short-trajectory RPE fix and
paper_comparison.mdrefresh. pcd_dogfooding --summary-jsonexports optionalrpe_trans_pct/rpe_rot_deg_per_m.- KITTI Odometry preparation script
evaluation/scripts/prepare_kitti_odometry_inputs.pygeneralized;setup_kitti_benchmark.shis a wrapper.
Saturated:
- CT-ICP: 5-cluster recipe structure mapped across 13 dataset/window combinations. Knob axes + seed-dependence saturated.
- LiTAMIN2: cluster T1 universal across 12 locally-available CT-ICP-comparable datasets (11/12 wins, 1/12 tied at noise floor).
- KISS-ICP / small_gicp / voxel_gicp: recipe pattern on 5 KITTI Odom seqs + 7 cross-dataset scenes. Seed-dependence verified (all 3 + LiTAMIN2 require GT seed for long-traj).
- NDT: T1 r05+i12 (
--ndt-resolution 0.5 --ndt-max-iterations 12) confirmed on KITTI Odom 5/5 full with sub-8cm seeded ATE. Remaining gap is throughput/implementation efficiency, not recipe universality on this benchmark. - A-LOAM: KITTI Odom 5/5 full transfer checked.
kitti_defaultis stable but drift-level; no universal sub-meter recipe. - F-LOAM: KITTI Odom 5/5 checked via seq 07 cluster + seq 00/02/05/08 transfer. Useful on seq 00/08, especially
denseon seq 00 andkitti_defaulton seq 08, but non-universal and drift-level. - LeGO-LOAM: KITTI Odom 5/5 checked via seq 07 cluster + seq 00/02/05/08 transfer. Strongest LOAM-family result on seq 02 and near-tie on seq 05, but non-universal and drift-level.
- MULLS: KITTI Odom seq 07 full cluster probe checked.
denseis the seq 07 accuracy/RPE winner, but all profiles are slow and drift-heavy; deprioritize held-out transfer.
Method-level production deployment recommendation:
- Seeded benchmark winners (KITTI Odom 5/5): NDT T1 r05+i12 (0.058-0.076 m), LiTAMIN2 T1 (0.65-0.75 m), small_gicp fast (0.68-0.98 m), voxel_gicp dense (0.94-1.05 m).
- No-seed deployment (production realism): CT-ICP cluster A only (12.69 m on seq 00; all others ≥87 m).
Priority order for next assistant:
- A (highest leverage): checkpoint the NDT + LOAM-family artifact/docs refresh before more sweeps. The current dirty set is large and internally consistent.
- B: resume cross-method universal recipe survey: apply LiTAMIN2 T1 / small_gicp fast / voxel_gicp dense to remaining local datasets (autoware_istanbul, hdl_400) to fill the cross-dataset matrix.
- C (broader, blocked on external data):
MulRan dcc01and Newer College ingestion. 2 pending dcc01 manifests already exist; data download required.
Do not spend the next turn on paper drafting, PR polishing, or speculative refactoring. The user has been explicit that this is OSS infrastructure work, not paper writing.
Working tree should be clean after the NDT seq 02/05/08 confirmation artifact refresh and LOAM-family KITTI Odom full transfer/cluster sweep checkpoint commit. Previous handoffs warned about a dirty worktree with mass untracked multimodal work; that state has since been committed. Treat git status as authoritative.
The branch is currently ahead of origin/wip/profile-expansion-refresh by some commits; verify with git status before pushing.
build/evaluation/pcd_dogfooding— LiDAR-only stable benchmark, 27 method families, shared--summary-jsoncontract.build/evaluation/multimodal_dogfooding— camera-aware sibling, 6 method families, same aggregate style and runner contract.
Each method emits:
status— lower-case enum fromdocs/status_taxonomy.md. Current emitting set:ok,skipped,timeout_budget. Reserved (not yet emitted):tracking_lost,init_failed,input_unsupported,metric_invalid,no_gt.ate_mrpe_trans_pct(optional; null when not computable)rpe_rot_deg_per_m(optional)framestime_msfpsnote
MethodResult::status is a std::string in C++ side; future detection logic should set it directly (e.g. result.status = "tracking_lost";) rather than introducing new boolean flags.
- Manifests:
experiments/*_matrix.json(336 active) - Pending manifests:
experiments/pending/*_matrix.json(200) - Aggregates:
experiments/results/*.json(336) - Family registry:
experiments/families.json— used by docs, not by the runner. - Generated docs:
docs/experiments.mddocs/decisions.mddocs/interfaces.mddocs/paper_comparison.mddocs/variant_analysis.mddocs/reproduction_status.md
Important: do not assume datasets mentioned in docs are locally present.
| Dataset | Path | Frames | Multimodal extras |
|---|---|---|---|
| HDL-400 native/reference-like | hdl_400_open_ct_lio_120 |
120 | imu.csv, frame_timestamps.csv |
| HDL-400 synthetic time (azimuth) | hdl_400_open_ct_lio_120_time_azimuth |
120 | imu.csv, frame_timestamps.csv |
| HDL-400 public ROS1 synthetic time | hdl_400_ros1_open_ct_lio_120_time_index |
120 | imu.csv, frame_timestamps.csv |
| KITTI Odometry seq 00 short | kitti_seq_00_108 |
108 | (LiDAR-only) |
| KITTI Odometry seq 00 full | kitti_seq_00_full |
4542 | (LiDAR-only) |
| KITTI Odometry seq 02 full | kitti_seq_02_full |
4661 | (LiDAR-only) |
| KITTI Odometry seq 05 full | kitti_seq_05_full |
2761 | (LiDAR-only) |
| KITTI Odometry seq 07 short | kitti_seq_07_108 |
108 | (LiDAR-only) |
| KITTI Odometry seq 07 full | kitti_seq_07_full |
1102 | (LiDAR-only) |
| KITTI Odometry seq 08 full | kitti_seq_08_full |
4071 | (LiDAR-only) |
| KITTI Raw 0009 short | kitti_raw_0009_200 |
200 | full multimodal extras |
| KITTI Raw 0009 full | kitti_raw_0009_full |
443 | full multimodal extras |
| KITTI Raw 0061 short | kitti_raw_0061_200 |
200 | full multimodal extras |
| KITTI Raw 0061 full | kitti_raw_0061_full |
703 | full multimodal extras |
| MCD KTH day-06 | mcd_kth_day_06_108 |
108 | frame_timestamps.csv |
| MCD NTU day-02 | mcd_ntu_day_02_108 |
108 | frame_timestamps.csv |
| MCD TUHH night-09 | mcd_tuhh_night_09_108 |
108 | frame_timestamps.csv |
| MulRan parkinglot 120 | mulran_parkinglot_120 |
120 | (LiDAR-only) |
| MulRan parkinglot full | mulran_parkinglot_full |
1177 | (LiDAR-only) |
GT CSVs for every dataset listed in 4.1. Verify by listing the directory before quoting paths.
- KITTI Odometry seq 01, 03, 04, 06, 09, 10 (only 00/02/05/07/08 are dogfooded)
- Istanbul local windows
- MulRan dcc01 / kaist / riverside (parkinglot only is dogfooded; dcc01 pending manifests exist)
- Newer College, Oxford Spires
This matters because:
- The next assistant must not claim runs on KITTI Odom seqs outside 00/02/05/07/08 without first ingesting those seqs.
- The 5 ingested KITTI Odom seqs have been exhaustively probed by CT-ICP; LiTAMIN2 and other LiDAR families have NOT yet been similarly probed.
aloam, balm2, clins, ct_icp, ct_lio, dlio, dlo, fast_lio2, fast_lio_slam, floam, genz_icp, gicp, hdl_graph_slam, isc_loam, kiss_icp, lego_loam, lins, lio_sam, litamin2, loam_livox, mulls, ndt, point_lio, rko_lio, small_gicp, suma, vgicp_slam, voxel_gicp, xicp
(genz_icp, rko_lio added 2026-06-02 later session — see §0.0.)
vins_fusion, okvis, orb_slam3, lvi_sam, fast_livo2, r2live
See experiments/families.json. Roughly:
- core (15):
kiss_icp,ct_icp,litamin2,small_gicp,voxel_gicp,ndt,gicp,fast_lio2,lio_sam,dlio,dlo,hdl_graph_slam,okvis,vins_fusion,fast_livo2 - extended (18): everything else, plus
orb_slam3,lvi_sam,r2livemarkedtimeout_prone.
Outside the current benchmark surface: vilens, relead, ct_icp_relead, scan_context, imu_preintegration, cube_lio_repro.
LiDAR-only coverage is already broad. The missing frontier is publication-grade reproduction evidence, not more method folders.
- Binary:
evaluation/src/multimodal_dogfooding.cpp - Windows:
kitti_raw_0009_200,kitti_raw_0009_full,kitti_raw_0061_200,kitti_raw_0061_full - Each run expects
sequence_dir,gt_csv,landmarks.csv,visual_observations.csv, camera intrinsics via CLI /camera_args.txt.
Fully practical under current study budget: okvis, vins_fusion, fast_livo2.
Practical-budget timeouts: orb_slam3, lvi_sam, r2live.
Canonical timeout budget (encoded in manifests via problem.variant_timeout_seconds):
- 200-frame windows: 120s
- Full windows: 300s
These align with smoke_200f_120s and practical_full_300s profiles in docs/budget_profiles.md.
evaluation/scripts/prepare_kitti_multimodal_inputs.pyevaluation/scripts/generate_kitti_visual_observations.pyevaluation/scripts/run_multimodal_study.pyevaluation/scripts/smoke_multimodal_fixture.sh
Per docs/reproduction_status.md and the stricter
promotion bar in docs/paper_ready_reproducibility.md:
| Method | Claim level | Why |
|---|---|---|
litamin2 |
indicative |
Short-window ATE, not full-sequence KITTI RPE study. |
ct_icp |
indicative |
Core close to paper formulation, but evaluation indirect. |
kiss_icp |
indicative |
Compact local-map pipeline, windowed reruns only. |
gicp |
ported |
Pre-KITTI paper; no standardized numeric target. |
ndt |
ported |
Pre-KITTI; modern NDT codebases differ materially. |
ct_lio |
ported |
Intentionally custom integration; no single paper-faithful target. |
The repo is not entitled to say "paper results reproduced" generally. The claim-level scheme makes that boundary explicit, while the paper-ready plan defines which methods can be promoted into a manuscript table.
LiTAMIN2 and CT-ICP are closer to reproducible paper-style evaluation because:
pcd_dogfoodingexports RPE.- Aggregates preserve RPE.
paper_comparison.mdshows repo-side RPE where available.
Canonical KITTI Raw 0009 reruns with RPE exist:
experiments/results/litamin2_kitti_raw_0009_matrix.json— adoptedfast_cov_half_threads(ATE 1.053 m, RPE trans 0.742 %, 34.16 FPS)experiments/results/ct_icp_kitti_raw_0009_matrix.json— adoptedfast_window(ATE 2.728 m, RPE trans 2.198 %, 49.27 FPS)
To move litamin2 and ct_icp from indicative toward approximately_reproduced:
- CT-ICP: full KITTI Odom 00/02/05/07/08 are now ingested and exhaustively probed. Per-seq best numbers (12.69 / 50.63 / 7.76 / 1.61 / 27.85 m ATE) are recipe-tuned, not paper-tuned. Direct paper-comparison requires either (a) committing to a single "paper-style" recipe and reporting per-seq deltas, or (b) reporting per-seq best with explicit recipe attribution.
- LiTAMIN2: same 5-cluster recipe-discovery approach has NOT yet been applied. This is the next high-leverage probe — analog of CT-ICP's cluster A/B/C/D/balanced structure may exist for LiTAMIN2.
- A (parallel): extend RPE-aware reruns to MCD/HDL-400 for both methods.
evaluation/scripts/prepare_kitti_odometry_inputs.py- Converts
sequences/<seq>/velodyne/*.bin+poses/<seq>.txtinto:dogfooding_results/kitti_seq_<seq>_<window>dogfooding_results/kitti_seq_<seq>_fullexperiments/reference_data/kitti_seq_<seq>_<window>_gt.csvexperiments/reference_data/kitti_seq_<seq>_full_gt.csv
evaluation/scripts/setup_kitti_benchmark.sh delegates to the Python script.
CT-ICP full KITTI Odom 00/02/05/07/08 manifests have been promoted to active (this session). Per-seq best results recorded; 5-cluster recipe structure documented in memory/ct_icp_kitti_full_per_seq_best.md. Cross-dataset cluster A transfer documented in memory/ct_icp_cluster_a_cross_dataset_transfer.md.
LiTAMIN2 full-sequence manifests remain in experiments/pending/ — runnable now that KITTI Odom data is ingested. This is the highest-leverage next probe.
tests/test_experiment_scripts.py contains a fake KITTI Odometry root test that verifies window export, full export, GT CSV generation, and contract compatibility of the preparation script.
- MulRan parkinglot: ingested. 8 indexed manifests covering
LiTAMIN2,CT-ICP,KISS-ICP,GICP× {120-frame, full}. - MulRan dcc01: 4 pending manifests in
experiments/pending/(*_mulran_dcc01_120_matrix.jsonfor ct_icp/gicp/kiss_icp/litamin2). Blocked: raw data not local. Requires MulRan official download form (https://forms.gle/EmUybUiGc8pR3r7Q6) perevaluation/scripts/SETUP_MULRAN_BENCHMARK.md. - Newer College: 2 pending manifests in
experiments/pending/(ct_icp_newer_college_01_short_120_matrix.json,litamin2_newer_college_01_short_120_matrix.json). Blocked on two fronts: (a) the manifests target sequence01_short_experiment, which is not present underdata/newer_college/; (b) the available localmath_harddata is a ROS2 bag (math_hard.bag+_rosbag2.db3) with topicos_node/lidar_packets(raw Ouster packets), not the flat PCD layout expected byprepare_newer_college_inputs.py. Extracting requires either downloading the flat-file release from https://ori-drs.github.io/newer-college-dataset/, or setting up ROS2 + Ouster driver to replay the bag and save PCDs.
MulRan dcc01 ingestion to complete the MulRan coverage. Reuse the existing parkinglot ingestion approach. Friendly format, broadens public-data evidence on the LiDAR-only surface. Pending the download form.
Newer College. LiDAR + IMU + camera; lets the repo claim multimodal coverage beyond KITTI Raw. Pending the flat-file download or a bag-extraction setup.
Oxford Spires. Newer, larger, more visually attractive, but higher ingestion cost than MulRan.
If the user says "do something useful without waiting for KITTI Odometry / MulRan dcc01 / Newer College data":
- Phase A reruns first (LiTAMIN2/CT-ICP on local MCD + HDL-400 with RPE) — cheapest claim-level improvement.
- Cross-method --no-gt-seed sweep on local KITTI Raw 0009 full — already done 2026-05-18, see
docs/dogfooding_methodology.md. Extending to other local datasets (MulRan parkinglot full, MCD KTH/NTU/TUHH) would cost almost nothing. - Then MulRan dcc01 when data lands — natural extension of partially-done work.
python3 -m unittest discover -s tests -p 'test_*.py' -v→ 14/14 passedcmake --build build --target pcd_dogfooding multimodal_dogfooding→ built clean- End-to-end runner smoke against
experiments/kiss_icp_kitti_raw_0009_matrix.json→ 3 variants completed, aggregate JSON emits new"status": "ok"taxonomy.
run_experiment_matrix.py— reuse-aggregate, timeout (now expects"timeout_budget"), RPE parsing.refresh_study_docs.pygenerate_reproduction_status.pygenerate_paper_comparison.pydata plumbingrun_multimodal_study.py- Synthetic multimodal fixture generator
- KITTI Odometry preparation script
- Full
ctest - Full Docker rebuild
- Any real KITTI Odometry full-sequence benchmark run
- Any new external dataset import beyond what is already on disk
Be precise about verification scope when reporting.
Generated docs already reflect canonical aggregate state at HEAD 5a96dec:
docs/interfaces.mddocs/experiments.mddocs/decisions.mddocs/paper_comparison.mddocs/variant_analysis.mddocs/reproduction_status.md— now includes claim-level legend.
To refresh all of them: python3 evaluation/scripts/refresh_study_docs.py.
This is the operational handoff. Default path: paper-ready reproducibility hardening (§00.52i), not new-method churn. Pick a single path and finish it before switching.
- Next action: add the next missing public/cross-domain check: M-GCLO public non-flat dataset validation, Quadric-LO public curved-object / non-urban dataset validation, or RF-LIO/ID-LIO public high-dynamic dataset validation. I-LOAM intensity on/off, KC-LO sigma schedule, M-GCLO ground factor, Quadric-LO plane fallback, and RF-LIO/ID-LIO synthetic dynamic stress plus M-GCLO synthetic non-flat stress plus Quadric-LO synthetic curved stress artifacts are already committed in §00.52j〜§00.52q.
- Extend the paper bundle:
docs/benchmarks/paper_ready_bundle.jsoncurrently freezes 4 methods. Grow it toward the final 8-12 method table from raw JSON. - Keep claims tiered: README may advertise breadth; paper/manuscript language should use only
T0/T1 methods from
docs/paper_ready_reproducibility.md. - Respect current user direction: 2D scan odometry is paused unless explicitly resumed.
- Verify docs after README/index edits:
python3 evaluation/scripts/validate_showcase.py --root .
- Resume new author-code-free 3D method implementation only after the paper-ready subset has a credible ablation table, or when the user explicitly asks for another "tugi".
- Keep the unit of work stable: module under
papers/<method>/, CMake integration,pcd_dogfoodingselector, focused unit tests, KITTI seq00/07 full artifacts, README leaderboard row,docs/methods.json, method README, and this PLAN.
- Commit + push pending work if user asks:
- IDC (
361a592), CSM-DT (3fc5be0), markdown hub (§00.59 files).
- IDC (
- Unify benchmarks:
Merge 8 methods into
cmake --build build --target scan_dogfooding bash evaluation/scripts/run_scan2d_benchmark.sh
docs/benchmarks/scan2d/{intel,fr079,mit,corridor}.json. - Verify CI path:
bash evaluation/scripts/smoke_scan2d_fixture.sh - Next 2D work: PL-ICP/MbICP local map or Karto-style map matcher.
- Docs: keep
docs/benchmarks/scan2d/README.mdas canonical; README top-level table stays a 1-screen summary linking to the hub.
- After README/index.html edits:
python3 evaluation/scripts/validate_showcase.py --root . - Demo path:
bash evaluation/scripts/demo_localization_zoo.sh
Blocked only by data.
- Obtain KITTI Odometry root containing at minimum:
sequences/00/velodyne+poses/00.txtsequences/07/velodyne+poses/07.txt
- Prep:
python3 evaluation/scripts/prepare_kitti_odometry_inputs.py \ --kitti-root <path> --sequence 00 --sequence 07 \ --window-size 108 --include-full
- Run 108-frame manifests first, then full-sequence manifests.
NCLT 600 honest negative (33% drift). Do not prioritize over 2D unless user redirects.
- Treat paper 50 MbICP as unstarted
- Resume 2D campaign without user direction
- Paper drafting / prose generation before the ablation table exists
- PR / branch cleanup unrelated to current task
- Broad refactor that touches the stable contract
- Force push / git config changes
- Reverting any pending manifest
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j"$(nproc)" --target pcd_dogfooding multimodal_dogfooding scan_dogfooding# CI-equivalent smoke (Intel 20f × 8 methods)
bash evaluation/scripts/smoke_scan2d_fixture.sh
# Refresh all committed 2D benchmarks
bash evaluation/scripts/run_scan2d_benchmark.sh
# Single fixture
./build/evaluation/scan_dogfooding \
evaluation/fixtures/intel_val_73 \
evaluation/fixtures/intel_val_73/gt.csv \
--methods rf2o,pl_icp,csm,kinematic_icp,psm,ndt_2d,idc \
--wheel-odom-from-gt \
--summary-json docs/benchmarks/scan2d/intel_val_73.jsonHub: docs/benchmarks/scan2d/README.md
python3 -m unittest discover -s tests -p 'test_*.py' -v
python3 evaluation/scripts/verify_environment.pypython3 evaluation/scripts/refresh_study_docs.py./build/evaluation/pcd_dogfooding <pcd_dir> <gt_csv> \
--methods litamin2,gicp,ndt \
--summary-json artifacts/tmp/smoke.json./build/evaluation/multimodal_dogfooding <sequence_dir> <gt_csv> \
--methods okvis,vins_fusion,fast_livo2 \
--landmarks-csv <sequence_dir>/landmarks.csv \
--visual-observations-csv <sequence_dir>/visual_observations.csv \
--summary-json artifacts/tmp/multimodal_smoke.jsonpython3 evaluation/scripts/prepare_kitti_multimodal_inputs.py --include-full
python3 evaluation/scripts/run_multimodal_study.py --include-full --method okvispython3 evaluation/scripts/prepare_kitti_odometry_inputs.py \
--kitti-root data/kitti_odometry \
--sequence 00 --sequence 07 \
--window-size 108 --include-fullUse repo-local scratch outputs when you do not want to disturb canonical repo docs / index:
python3 evaluation/scripts/run_experiment_matrix.py \
--manifest experiments/<something>.json \
--output-dir artifacts/tmp/localization_zoo_results \
--docs-dir artifacts/tmp/localization_zoo_docs- 2026-06-10 更新: アクティブ価値は 2D scan odometry zoo (8 法 × 4 fixture, honest leaderboard) と from-paper 50 本の蓄積。§00.6c–§00.60 が運用 handoff。
- 3D 側は依然 valuable だが、PG-LIO honest negative と KITTI full データ欠如で 新規 3D 論文より 2D 拡張 (local map / Karto-style matcher) を優先。
- The cheapest claim-level upgrade for 3D remains local MCD/HDL-400 reruns with RPE; strongest is full KITTI Odometry (blocked).
- Keep the stable summary contract small; new failure modes flow through
MethodResult::statusstrings. - Do not oversell reproduction status. The taxonomy exists precisely so the repo can be honest.
- 2D GT proxy 限界: Bonn dataset odometry は centimeter truth ではない — drift は相対比較用。