Localization Zoo

C++ implementations, derived variants, and compact baselines for localization papers

103 methods · 73 paper reimplementations · 42 papers with no public author code · one C++ API · honest KITTI benchmarks

_{KITTI seq07 odometry, no GT seed — current promoted trajectories on one shared drift-colored scale.}

↪ Open the interactive benchmark and method explorer

Try it in one command — no build, no dataset; runs the quick synthetic and committed-data comparison, then drops report.html and its reproducibility manifest.json into ./zoo-demo/:

docker run --rm -v "$PWD/zoo-demo:/out" ghcr.io/rsasaki0109/localization_zoo:latest

Promoted LiDAR odometry v10

direction_consistent_rotation_v10 combines the motion-guard KISS frontend with official KISS-ICP 1.3.0. Ground truth is used only for scoring.

Public evaluation	v10 ATE	v10 translational RPE	Conservative rate	Result
Fresh KITTI Raw 0023 (474 frames)	5.494 m	1.622%	12.38 FPS	ATE −2.00% vs v6; RPE +0.146%
Same-input KITTI Odometry 04 (271 frames)	1.145 m	0.354%	10.67 FPS	Best ATE and translational RPE in the frozen six-method table
External Boreas summer drive (600 frames)	0.663 m	0.966%	3.02 FPS	ATE −0.407% and RPE −0.100% vs v6; dense-sensor runtime limitation

The frozen comparison also includes v6, MOLA-LO, MAD-ICP, official KISS-ICP, and CT-ICP. See the frozen manifest, aggregate, and protocol. v10 is not universally best; its dual frontend is also not real-time on dense Boreas scans.

Re-score the frozen trajectories and verify all hashes from the external SSD:

export LOCALIZATION_ZOO_DATA_ROOT=/media/external/loc_zoo
python3 evaluation/scripts/reproduce_lidar_v10.py

On PowerShell, set $env:LOCALIZATION_ZOO_DATA_ROOT before running the command.

Why Localization Zoo?

Paper-oriented C++ implementations share one API, ROS 2 wrappers, tests, and evaluation tools—even when no reusable author code exists.

Pure C++: ROS-independent core libraries for research, education, and embedded use
ROS 2 Humble: Ready-to-run nodes via ros2 run
Built-in benchmarking: Compare methods immediately after build
Degeneracy-aware methods: Includes RELEAD and X-ICP for constrained environments
Claim tiers: Separate faithful reproduction, mechanism evidence, adapters, and compact baselines

Paper-Ready Claim Boundary

The catalog is broader than the manuscript-grade evidence set:

Catalog: all 103 methods, including derived variants and compact baselines.
Paper-ready subset: methods satisfying the tier and ablation criteria, frozen in paper_ready_bundle.json.
Frozen paper-ready bundle (8 methods): I-LOAM and KC-LO (T0 candidates); LiDAR-IBA and TrICP-LO (T1); DegenSense and D2-LIO (competitive no-IMU KITTI fallback rows, not full LIO claims); M-GCLO and Quadric-LO (T1+). Manifest: paper_ready_bundle.json. Other ports remain adapter/mechanism evidence until their sensor-data gaps close.

Leaderboard — odometry, RPE [drift %/100 m], lower is better

KITTI Odometry full sequences, ranked by relative pose error (the seed-independent local-accuracy metric). ATE in parens is unbounded drift over the run. Full matrix: explorer.

Method	Seq 00	Seq 02	Seq 05	Seq 07	Seq 08
	4541 fr	4661 fr	2761 fr	1101 fr	4071 fr
LeGO-LOAM	0.84% _{(12 m)}	0.88% _{(42 m)}	0.52% _{(5 m)}	0.53% _{(3 m)}	1.37% _{(19 m)}
A-LOAM	0.90% _{(19 m)}	0.93% _{(51 m)}	0.51% _{(5 m)}	0.61% _{(3 m)}	1.39% _{(19 m)}
KISS-ICP	0.86% _{(21 m)}	0.94% _{(39 m)}	0.62% _{(6 m)}	0.61% _{(2 m)}	1.34% _{(19 m)}
F-LOAM	0.99% _{(9 m)}	0.95% _{(54 m)}	0.54% _{(6 m)}	0.59% _{(3 m)}	1.37% _{(17 m)}
SuMa	–	–	–	0.94% _{(4 m)}	–
CT-ICP	1.97% _{(19 m)}	2.64% _{(67 m)}	1.09% _{(11 m)}	1.17% _{(3 m)}	1.91% _{(31 m)}
MULLS	–	–	–	2.64% _{(8 m)}	–

Best variant per cell (docs/experiments.md). KISS-ICP / LOAM ~0.5–1.4% drift is competitive — their large ATE is honest drift, not a broken port.

No GT-seeded methods here. NDT / LiTAMIN2 / GICP use the ground-truth pose as the per-frame initial guess, so their ATE is seed adherence, not tracking — and the ranking even inverts: NDT's 0.02 m comes from not registering (--no-gt-seed → 87% RPE, the worst tracker), while GICP's larger ATE is real registration. They need a GT prior and aren't standalone odometry, so they are not ranked.

From-paper reimplementations (no public reference code) — KITTI full

Details and negative results: Reproducibility Report.

KITTI full-sequence odometry with first-pose anchor and --no-gt-seed. RPE is drift %/100 m; ATE is shown in parentheses. LIO rows use a no-IMU fallback.

This is a shared-harness catalog, not an endorsement of paper fidelity. Manuscript evidence is limited to the frozen bundle.

Method	Seq 00 (4541 fr)	Seq 07 (1101 fr)	Paper
DegenSense	0.811% _{(12 m)}	0.558% _{(1 m)}	arXiv:2412.07513
D2-LIO	0.814% _{(11 m)}	0.541% _{(1 m)}	arXiv:2508.14355
M-GCLO	0.835% _{(19 m)}	0.671% _{(2 m)}	ISPRS Ann. 2024
KC-LO	0.837% _{(13 m)}	0.510% _{(1 m)}	ECCV 2004
LiDAR-IBA	0.841% _{(11 m)}	0.633% _{(1 m)}	arXiv:2602.06380
LODESTAR	0.848% _{(7 m)}	0.598% _{(1 m)}	arXiv:2511.09142
Terrain-RBF-LIO	0.849% _{(8 m)}	0.587% _{(1 m)}	arXiv:2509.26222
DALI-SLAM	0.849% _{(8 m)}	0.600% _{(1 m)}	ISPRS JPRS 2025
DAMM-LOAM	0.851% _{(7 m)}	0.598% _{(1 m)}	arXiv:2510.13287
CUBE-LIO	0.851% _{(9 m)}	0.608% _{(1 m)}	ROBOMECH 2026
Intensity-Flow	0.856% _{(8 m)}	0.616% _{(1 m)}	Measurement 2026
Quadric-LO	0.867% _{(15 m)}	0.598% _{(2 m)}	arXiv:2304.14190
Adaptive-ICP	0.870% _{(11 m)}	0.569% _{(1 m)}	arXiv:2509.22058
MCC-LO	0.892% _{(13 m)}	0.611% _{(2 m)}	PLOS ONE 2018
I-LOAM	0.899% _{(13 m)}	0.575% _{(2 m)}	UR 2020
Mesh-LOAM	0.901% _{(13 m)}	0.616% _{(1 m)}	IEEE T-IV 2024
NHC-LIO	0.902% _{(18 m)}	0.608% _{(3 m)}	IEEE Sens. J. 2023
SVN-ICP	0.912% _{(14 m)}	0.607% _{(3 m)}	arXiv:2509.08069
ICPSC-LO	0.912% _{(19 m)}	0.660% _{(4 m)}	JAG 2023
VLOM	0.914% _{(10 m)}	0.605% _{(3 m)}	arXiv:2304.08978
TrICP-LO	0.931% _{(10 m)}	0.662% _{(2 m)}	IVC 2005
GMM-LO	0.941% _{(14 m)}	0.657% _{(1 m)}	arXiv:1807.02587
MCGICP-LO	0.940% _{(20 m)}	0.774% _{(5 m)}	RAS 2017
Student-T-LO	0.952% _{(15 m)}	0.696% _{(2 m)}	PMC11314997 2024
Small-but-Mighty	0.961% _{(15 m)}	0.897% _{(3 m)}	Remote Sens. 2025
GNC-LO	0.986% _{(18 m)}	0.722% _{(2 m)}	arXiv:1909.08605
IMLS-SLAM	1.000% _{(18 m)}	0.700% _{(3 m)}	ICRA 2018
CT-VoxelMap	1.046% _{(21 m)}	0.800% _{(3 m)}	arXiv:2604.03747
OPL-LVIO	1.050% _{(15 m)}	0.902% _{(4 m)}	Remote Sens. 2022
AD-VLO	1.052% _{(13 m)}	0.939% _{(3 m)}	ICRA 2019
TC-MVLO	1.054% _{(11 m)}	0.939% _{(3 m)}	ISR 2022
TC-LVGF	1.055% _{(12 m)}	0.941% _{(4 m)}	ROBIO 2023
TC-VLO	1.060% _{(12 m)}	0.925% _{(4 m)}	IV 2019
BIEVR-LIO	1.063% _{(25 m)}	0.873% _{(4 m)}	arXiv:2604.14421
V-LOAM2015	1.066% _{(14 m)}	0.910% _{(4 m)}	ICRA 2015
Vibration-LIO	1.082% _{(15 m)}	0.781% _{(3 m)}	arXiv:2507.04311
ID-LIO	1.111% _{(15 m)}	0.999% _{(5 m)}	Sensors 2023
ELO	1.124% _{(23 m)}	0.981% _{(4 m)}	IEEE RA-L 2021
UA-LIO	1.132% _{(33 m)}	0.967% _{(3 m)}	IEEE TIM 2025
DiLO	1.200% _{(39 m)}	1.533% _{(7 m)}	ETRI J. 2021
PCR-DAT	1.239% _{(11 m)}	1.040% _{(4 m)}	ISR 2024
RF-LIO	1.351% _{(23 m)}	1.272% _{(5 m)}	IROS 2021
Spectral-LO	2.901% _{(67 m)}	3.939% _{(27 m)}	arXiv:2005.02042
KISS-ICP (same profile, ref)	0.872% _{(12 m)}	0.618% _{(2 m)}	—
CT-ICP (same profile, ref)	2.577% _{(17 m)}	2.500% _{(4 m)}	—

Mechanism evidence and limitations are summarized below; raw artifacts are linked.

HDL-400 activates IMU paths for D2-LIO, DegenSense, ID-LIO, and RF-LIO, with small RPE deltas on this window (artifact). LiDAR-IBA IMU residuals are not yet wired into pcd_dogfooding.

NCLT confirms IMU-gated paths and improves DegenSense ATE from 0.24 m to 0.16 m; poor KISS-ICP sanity makes this mechanism evidence only (artifact).

M-GCLO: disabling ground factors worsens ATE and rotational drift on KITTI (ablation) and synthetic rolling ground (stress test).

On hilly KITTI seq08, ground-off keeps RPE similar but worsens ATE by 149% (artifact).

MulRan no-GT-seed odometry diverges; the GT-seeded result is mechanism evidence only (artifact).

KC-LO: best seq07 drift here, but only ~1.4–3.1 FPS (ablation). LiDAR-IBA: bundle adjustment slightly lowers ATE but worsens RPE and throughput (ablation). Quadric-LO: plane fallback is rare on highway data (ablation; curved-object stress).

On KITTI seq02, disabling the rare fallback worsens RPE by 55% and ATE by 84% (artifact).

On geometry-rich, IMU-free KITTI, many robust and multimodal mechanisms become secondary to the point-to-plane core. Dynamic-scene stress still favors RF-LIO over ID-LIO (artifact).

On KITTI seq05, RF-LIO trails ID-LIO and KISS-ICP (artifact).

Full ablations and negative results: benchmark artifacts.

Does LiDAR intensity actually help? — I-LOAM ablation

I-LOAM intensity on/off results with mapping disabled:

Sequence	Geometric baseline (intensity off)	I-LOAM (intensity on)	Δ drift
Seq 00	3.186% _{(76.0 m)}	2.606% _{(49.4 m)}	−18.2%
Seq 07	3.806% _{(18.5 m)}	3.055% _{(15.1 m)}	−19.7%

Reflectance cuts drift by 18–20% in this paired KITTI ablation. Raw artifacts: seq00 on, seq00 off, seq07 on, seq07 off, and the paired summary i_loam_intensity_ablation.json.

Trajectory gallery — KITTI seq07, current promoted trajectories

The hero shows KITTI seq07 full trajectories on one drift-colored scale. Regenerate it or use the generic plotter.

Scope Note

Methods are labeled paper reimplementation, derived variant, or compact baseline. A benchmark result does not by itself prove paper fidelity; see reproduction status and paper-ready criteria.

Experiment-Driven Development

The stable benchmark core is separated from variants in experiments/. See experiments, decisions, and interfaces.

Quick checks (after clone)

# build + synthetic benchmark + quick real-data fixture suite
bash evaluation/scripts/demo_localization_zoo.sh

Output is written under experiments/results/runs/demo_localization_zoo/. After the first successful run, compare the broader starter set without rebuilding:

bash evaluation/scripts/demo_localization_zoo.sh --skip-build --profile broad

CI-equivalent smoke checks:

bash evaluation/scripts/smoke_ci_fixture.sh
bash evaluation/scripts/smoke_multimodal_fixture.sh

More workflows: evaluation/README.md. Activation and retention definitions: docs/activation_metrics.md.

Benchmark

Compare results only within the same dataset, window, initialization policy, and runtime profile.

Current published groups:

Group	Role	Ranking policy	Source
KITTI Odometry seq00 full	Odometry full sequence	Ranked within seq00 by translational RPE [%/100 m]	`docs/benchmarks/paper_ready_bundle.json`
KITTI Odometry seq07 full	Odometry full sequence	Ranked within seq07 by translational RPE [%/100 m]	`docs/benchmarks/paper_ready_bundle.json`
KITTI Odometry seq07 108-frame smoke	Regression smoke	Unranked; exact frame-ID association check only	`docs/benchmarks/latest/results.json`
Autoware Istanbul 108-frame snapshot	GT-seeded scan-to-map references	Reference-only; not ranked against odometry	`docs/benchmarks/latest/results.json`

Detailed rows and provenance are in the interactive explorer, interfaces, and methodology. KITTI preparation example:

python3 evaluation/scripts/prepare_kitti_odometry_inputs.py \
  --kitti-root /path/to/data_odometry --sequence 00 --sequence 07 \
  --window-size 108 --include-full

Synthetic Urban (30 frames)

Method         ATE [m]     FPS
─────────────────────────────────
CT-ICP         0.124       0.1   << best accuracy
A-LOAM         2.059       4.6   << balanced baseline
X-ICP          16.634      5.9
LiTAMIN2       31.155      2.6

Reproduce with ./synthetic_benchmark. No external dataset is required.

2D Laser Scan Odometry

Eight planar matchers (papers 43–50) plus Karto-Matcher (map-based CSM extension) in scan_dogfooding. Full leaderboard, fixtures, and reproduction steps: docs/benchmarks/scan2d/README.md.

Method	Intel val	fr079 val	MIT val	Synth corridor
	73 fr / 378 m	384 fr / 373 m	33 fr / 267 m	120 fr / 9.5 m
RF2O	14.3%	15.4%	27.6%	1.3%
Karto-Matcher	14.0%	13.7%	28.1%	30.5%
CSM	14.0%	13.7%	28.1%	30.5%
NDT-2D	14.9%	14.4%	27.8%	0.8%
IDC	15.3%	27.7%	29.5%	42.6%
MbICP	14.5%	15.4%	27.5%	0.05%
PL-ICP	15.0%	14.1%	27.2%	0.01%
Kinematic-ICP	18.4%	18.9%	23.4%	83.8%
PSM	21.8%	13.9%	27.9%	11.6%

cmake --build build --target scan_dogfooding
bash evaluation/scripts/run_scan2d_benchmark.sh

Setup: evaluation/scripts/SETUP_2D_SCAN_BENCHMARK.md.

Hard Point Cloud Localization

Hard Point Cloud Localization Dataset results; full protocol and coverage are in the benchmark README.

Full-trajectory indoor_easy_01 vs indoor_hard_01, six unchanged runner defaults:

Method	Easy ATE-XY (m)	Hard ATE-XY (m)	Easy RPE-XY (m/f)	Hard RPE-XY (m/f)
KISS keyframe	8.763	25.307	0.076	0.124
LiTAMIN2	10.815	17.875	0.358	0.611
CT-ICP	18.767	16.795	0.284	0.350
X-ICP	10.037	11.733	0.320	0.353
DegenSense + IMU	13.201	617.008	0.567	8.319
DegenSense, no IMU	11.759	11.729	2.736	3.679

CT-ICP's lower hard ATE is misleading: its estimated path is 468.38 m versus 115.49 m ground truth. Official BIEVR-LIO:

Sequence	ATE-XY (m)	RPE-XY (m/f)	Estimated / GT path (m)
`indoor_easy_01`	0.422	0.0033	76.26 / 77.29
`indoor_hard_01`	8,882.579	21.947	28,754.33 / 114.71

BIEVR-LIO excels on easy but diverges on hard. Fixed-map NDT also false-locks on all three kidnap sequences, showing that stable refinement alone does not prove a correct global lock.

python3 evaluation/scripts/run_hard_pcl_odometry_benchmark.py \
  --pcd-dir dogfooding_results/hard_pcl_localization/indoor_easy_01 \
  --gt-csv experiments/reference_data/hard_pcl_indoor_easy_01_gt.csv \
  --output-dir experiments/results/hard_pcl_localization/indoor_easy_01/full --jobs 6

Setup: evaluation/scripts/SETUP_HARD_PCL_LOCALIZATION_BENCHMARK.md. Full results tree: experiments/results/hard_pcl_localization/.

Implementations

Point Cloud Registration

Paper	Venue	Key Idea	Reference
LiTAMIN2	ICRA 2021	KL-divergence ICP with aggressive point reduction for faster registration	arXiv
GICP	RSS 2009	Plane-to-plane ICP with local covariance modeling and Mahalanobis distance	Paper
Voxel-GICP	RA-L 2021	GICP accelerated with voxel representatives and voxel-level covariance	Paper
small_gicp	Derived	Compact GICP with voxel downsampling and capped correspondences	GitHub
VGICP-SLAM	Derived	Voxel-GICP front-end with Scan Context and loop-graph back-end	-
NDT	IROS 2003	NDT-style registration against voxel Gaussian models with a compact optimizer	Paper
KISS-ICP	RA-L 2023	Compact KISS-ICP-style pipeline with voxel hashing, adaptive thresholds, and robust ICP	Paper
A-LOAM	RSS 2014	Curvature-based feature extraction with a three-stage odom-to-map pipeline	GitHub (ROS1)
F-LOAM	Derived	Lightweight LOAM pipeline with input thinning and sparse map updates	-
ISC-LOAM	Derived	Lightweight LOAM with intensity descriptors and F-LOAM/GICP loop graph	-
LOAM Livox	Derived	LOAM variant for solid-state LiDAR using pseudo scan lines from azimuth sectors	Reference
LeGO-LOAM	IROS 2018	Ground-aware feature extraction for vehicle-oriented LOAM	Paper
MULLS	Derived	Multi-metric scan-to-map alignment using edge, plane, and point residuals	-
BALM2	T-RO 2022	Local bundle adjustment over recent keyframes with line and plane residuals	arXiv
SuMa	RSS 2018	Dense surfel-based point-to-plane odometry from range-image maps	GitHub
DLO	Derived	Direct keyframe odometry that aligns dense scans to a local GICP map	GitHub
HDL Graph SLAM	Derived	NDT front-end with floor priors, Scan Context, and GICP loop closures	GitHub
CT-ICP	ICRA 2022	Continuous-time registration with two poses per frame and SLERP motion compensation	GitHub (ROS1)
X-ICP	T-RO 2024	Constrained ICP using Hessian SVD to classify observable directions	arXiv

Degeneracy-Aware Methods

Paper	Venue	Key Idea	Reference
RELEAD	ICRA 2024	Constrained ESIKF with projection-based suppression along degenerate directions	arXiv
CT-ICP + RELEAD	Hybrid	Continuous-time CT-ICP interpolation combined with RELEAD degeneracy handling	-

Foundations

Paper	Venue	Key Idea	Reference
IMU Preintegration	T-RO 2017	Preintegration on SO(3) with first-order bias correction for LIO pipelines	Paper

LIO / Continuous-Time Fusion

Paper	Venue	Key Idea	Reference
CT-LIO	Hybrid	Lightweight LIO that combines CT-ICP interpolation with IMU preintegration constraints	CLINS
DLIO	ICRA 2023	Compact DLIO-style LIO built on DLO-style registration plus IMU preintegration	GitHub
LINS	Derived	Lightweight LiDAR-inertial estimator with iterated filtering and point-to-plane updates	-
Point-LIO	Derived	Compact direct LIO with raw-point planarity correspondences and iterated filtering	-
CLINS	Derived	Sequence pipeline version of CT-LIO-style continuous-time registration	-
VILENS	Derived	Compact visual-lidar-inertial smoother with Point-LIO-style local maps and reprojection fusion	-
LIO-SAM	IROS 2020	Lightweight pose graph with A-LOAM front-end, Scan Context, GICP, and IMU rotation priors	Paper
LVI-SAM	Derived	Compact visual-lidar-inertial SLAM built on top of a LIO-SAM-style pose graph	-
VINS-Fusion	Derived	Compact visual-inertial odometry with landmark reprojection and IMU preintegration	-
OKVIS	Derived	Fixed-window VIO with landmark reprojection and IMU preintegration	-
ORB-SLAM3	Derived	Compact visual-inertial SLAM with keyframe graph and overlap-based loop closure	-
FAST-LIO2	T-RO 2022	Lightweight direct LIO with raw-point scan-to-map alignment and IMU prediction	Paper
FAST-LIVO2	Derived	Compact local visual-lidar-inertial odometry built on FAST-LIO2 poses and reprojection residuals	-
R2LIVE	Derived	Compact visual-lidar-inertial SLAM combining FAST-LIO2 odometry and visual landmark factors	-
FAST-LIO-SLAM	Derived	Lightweight graph SLAM with FAST-LIO2 front-end, Scan Context, and GICP loop closures	-

IMU Dead Reckoning / Aided INS

IMU-only methods (no LiDAR registration); evaluated on NCLT 2013-01-10 and KITTI Raw drive 0009 with imu.csv inputs, not on the KITTI LiDAR leaderboard.

Paper	Venue	Key Idea	Reference
IMU Dead Reckoning	Baseline	Strapdown INS lower-bound for the LIO family, with opt-in classical aids (ZUPT / NHC / attitude leveling / ZARU), a motion-gated static-init quality gate, and a 15-state error-state Kalman filter (`--imu-dr-eskf`) mode	-
OdoNet	IEEE Sens. J. 2022	1D-CNN pseudo-odometer speed aiding for strapdown INS with NHC/ZUPT	Paper
NN-ZUPT	Meas. Sci. Technol. 2023	CNN zero-velocity detection driving ZUPT corrections for vehicle INS	Paper
NHC-Net	GPS Solutions 2023	Motion-state CNN with adaptive non-holonomic constraints for vehicle dead reckoning	-

IMU Motion & Health SDK

imu_motion_health is a LiDAR-free streaming C++ SDK and CSV replay CLI for 6-axis IMUs. It diagnoses startup calibration, bias, timestamp gaps, non-finite values, and saturation; classifies stationary, moving, impact, fall/tilt, and vibration states; and emits JSON diagnostics plus short-term relative attitude, velocity, and position. Its standalone build only requires Eigen3 and GTest, and the included deterministic demo needs no sensor download.

Place Recognition / Loop Closure

Paper	Venue	Key Idea	Reference
Scan Context	IROS 2018	Lightweight place recognition with polar ring-sector descriptors and yaw-shift search	Paper

Quick Start

Docker (no local dependencies)

docker run --rm -v "$PWD/zoo-demo:/out" ghcr.io/rsasaki0109/localization_zoo:latest

Native build

sudo apt install libeigen3-dev libpcl-dev libopencv-dev libceres-dev libgtest-dev
python3 -m venv .venv
. .venv/bin/activate
python3 -m pip install -r requirements-lock.txt
bash evaluation/scripts/demo_localization_zoo.sh

Manual build and test:

cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j"$(nproc)"
ctest --test-dir build --output-on-failure

./build/evaluation/synthetic_benchmark

Python bindings (experimental)

Experimental pybind11 bindings cover KISS-ICP, NDT, GICP, and LiTAMIN2:

cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_PYTHON_BINDINGS=ON
cmake --build build -j"$(nproc)" --target localization_zoo_py
python3 python/example_synthetic.py

import sys; sys.path.insert(0, "build/python")
import localization_zoo as lz

odom = lz.KissICP()
for scan in scans:          # Nx3 float64 numpy arrays
    odom.register_frame(scan)
print(odom.pose)            # 4x4 world-frame pose

See python/README.md for the full API.

ROS 2

cd ros2 && colcon build
source install/setup.bash

ros2 run localization_zoo_ros litamin2_node
ros2 launch localization_zoo_ros play_rosbag.launch.py \
  bag_path:=/path/to/bag points_topic:=/velodyne_points

Evaluation

python3 evaluation/scripts/benchmark.py \
  --gt gt_poses.txt \
  --est LiTAMIN2:litamin2_poses.txt A-LOAM:aloam_poses.txt \
  --output_dir results/

Architecture

localization_zoo/
├── common/      # Shared Eigen/PCL utilities
├── papers/      # One self-contained dir per method (headers, sources, tests)
├── evaluation/  # Benchmark and evaluation tools
├── ros2/        # ROS 2 Humble wrappers
└── .github/workflows/ci.yml

Each papers/*/ directory is self-contained; core libraries are ROS-independent.

Degeneracy Detection Demo

RELEAD and X-ICP report underconstrained geometry:

=== Tunnel Environment ===
Has degeneracy: yes
Degenerate translation dirs: 1
  dir: [1, 0, 0]              # x direction (tunnel axis) is degenerate

=== Normal Environment ===
Has degeneracy: no             # walls and ground constrain all directions

ROS 2 publishes the status on /degeneracy.

Adding a New Paper

mkdir -p papers/your_method/{include/your_method,src,test}
# 1. Write headers, sources, and tests
# 2. Add CMakeLists.txt
# 3. Add add_subdirectory to the top-level CMakeLists.txt
# 4. Run ctest and keep the full suite passing

See CONTRIBUTING.md for the checklist, or open a paper request.

Dependencies

Library	Version	Purpose
Eigen3	>= 3.3	Linear algebra
PCL	>= 1.10	Point cloud processing
Ceres Solver	>= 2.0	Nonlinear optimization
GTest	>= 1.11	Unit testing
OpenCV	>= 4.0	I/O utilities

Contributing

Contributions are welcome — new paper reimplementations, bug fixes, and benchmark improvements. See CONTRIBUTING.md for scope, the honesty policy, and the per-paper checklist.

Citation

If you use this software or its benchmarks, please cite it (see CITATION.cff) or use GitHub's "Cite this repository" button.

@software{sasaki_localization_zoo,
  author  = {Sasaki, Ryohei},
  title   = {Localization Zoo: from-paper {C++} reimplementations of {LiDAR}
             localization and odometry methods, honestly benchmarked},
  url     = {https://github.com/rsasaki0109/localization_zoo},
  version = {1.0.0},
  year    = {2026}
}

License

MIT

Name		Name	Last commit message	Last commit date
.github		.github
common		common
docs		docs
evaluation		evaluation
experiments		experiments
papers		papers
python		python
ros2/localization_zoo_ros		ros2/localization_zoo_ros
tests		tests
third_party		third_party
.dockerignore		.dockerignore
.gitignore		.gitignore
CITATION.cff		CITATION.cff
CMakeLists.txt		CMakeLists.txt
CODEX_TASK.md		CODEX_TASK.md
CODEX_TASK_2_method_expansion.md		CODEX_TASK_2_method_expansion.md
CODEX_TASK_3_test_verification.md		CODEX_TASK_3_test_verification.md
CODEX_TASK_4_hard_pcl_localization.md		CODEX_TASK_4_hard_pcl_localization.md
CONTRIBUTING.md		CONTRIBUTING.md
Dockerfile		Dockerfile
Doxyfile		Doxyfile
LICENSE		LICENSE
PLAN.md		PLAN.md
README.md		README.md
requirements-lock.txt		requirements-lock.txt
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation

Localization Zoo

Promoted LiDAR odometry v10

Why Localization Zoo?

Paper-Ready Claim Boundary

Leaderboard — odometry, RPE [drift %/100 m], lower is better

From-paper reimplementations (no public reference code) — KITTI full

Does LiDAR intensity actually help? — I-LOAM ablation

Trajectory gallery — KITTI seq07, current promoted trajectories

Scope Note

Experiment-Driven Development

Quick checks (after clone)

Benchmark

Synthetic Urban (30 frames)

2D Laser Scan Odometry

Hard Point Cloud Localization

Implementations

Point Cloud Registration

Degeneracy-Aware Methods

Foundations

LIO / Continuous-Time Fusion

IMU Dead Reckoning / Aided INS

IMU Motion & Health SDK

Place Recognition / Loop Closure

Quick Start

Docker (no local dependencies)

Native build

Python bindings (experimental)

ROS 2

Evaluation

Architecture

Degeneracy Detection Demo

Adding a New Paper

Dependencies

Contributing

Citation

License

About

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages