Skip to content

DATA.md

Latest commit

 

History

History
65 lines (49 loc) · 2.62 KB

DATA.md

File metadata and controls

65 lines (49 loc) · 2.62 KB
Raw

Data preparation

OV-MAP reads datasets from a data/ directory and writes results to output/ at the repository root. Both are untracked by git — create them as local directories or symlinks to a large-storage location, e.g.:

ln -s /path/to/storage/OVMap/data   data
ln -s /path/to/storage/OVMap/output output

ScanNet / ScanNet200

Request access to and download ScanNet from the official repository: https://github.com/ScanNet/ScanNet (you must agree to the ScanNet Terms of Use). The ScanNet200 label set is derived from the same scans. The validation split used in the paper (312 scenes) is listed in scannet-preprocess/meta_data/scannetv2_val.txt.

Expected layout

data/scannetv2/
├── input/
│   ├── scannetv2_images/val/<scene>/{color,depth,pose,intrinsics,label}/   # posed RGB-D frames
│   ├── scannetv2_pcds/val/<scene>.ply, <scene>.pth                         # scene point clouds
│   └── mesh_segmentation/0.05_20/<scene>_vh_clean_2.0.050000.segs.json     # mesh over-segmentation
└── preprocessed_data/scannetv2_200_openmask3d/
    ├── label_database.yaml
    └── instance_gt/validation/<scene>.txt                                  # evaluation ground truth

Preprocessing steps

  1. Posed RGB-D frames + label images (from raw .sens scans):

    export SCANNET_GIT_DIR=/path/to/ScanNet
bash scripts/prepare_data.sh

    This runs scannet-preprocess/prepare_2d_data/prepare_2d_data.py. Point-cloud preprocessing (preprocess_scannet.py) and the OpenMask3D ground-truth preprocessing are included as commented steps in the same script — enable them as needed. These tools are adapted from the ScanNet toolkit and Pointcept (see the README acknowledgements).

  2. Mesh over-segmentation at kThresh = 0.05 (required by the dominant-voting step):

    ScanNet ships only the default 0.01 segmentation, so the 0.05 files must be generated with ScanNet's segmentator tool (build it from the ScanNet repo), then:

    SEGMENTATOR_BIN=/path/to/segmentator \
SCANNET_SCANS_DIR=/path/to/ScanNet/scans \
bash scripts/generate_mesh_segmentation.sh

    This writes data/scannetv2/input/mesh_segmentation/0.05_20/. The parameters are configured in ovmap/config.py (MESH_SEG_KTHRESH, MESH_SEG_SEGMIN).

Real-world scenes

The real-world experiments (Fig. 7) use RGB-D captured by the authors; this data is not publicly released. The ScanNet200 evaluation above reproduces the main quantitative results (Tables I and III).