Skip to content

datasets.md

Latest commit

 

History

History
116 lines (88 loc) · 3.3 KB

datasets.md

File metadata and controls

116 lines (88 loc) · 3.3 KB

External Datasets

PointCloud-GeoLab keeps large datasets out of git. The repository contains small synthetic demo data for CI, while real data should live under data/external/ or another local path passed to the examples.

Supported Sources

Dataset Official Source Local Use Notes
Stanford Bunny / Armadillo Stanford 3D Scanning Repository registration, reconstruction, OBB/PCA Bunny range data and reconstruction are small enough for laptop demos; Armadillo is larger.
KITTI Velodyne KITTI Raw Data LiDAR ground removal and clustering KITTI stores Velodyne frames as repeated float32 x y z intensity tuples in .bin.
ModelNet10/40 Princeton ModelNet download page primitive/PCA demos and optional ML ModelNet meshes are OFF files; convert a few samples to point clouds before running examples.

Recommended Layout

data/external/
  stanford/
    bunny/
      bunny.ply
    bunny_pair/
      bunny_source.ply
      bunny_target.ply
    armadillo/
      Armadillo.ply
  kitti/
    velodyne/
      000000.bin
  modelnet_small/
    chair.off
    sample.xyz
  manifest.json

Do not commit these files. Commit only scripts, docs, and tiny deterministic fixtures.

Preparation Commands

Print the expected layout:

python scripts/prepare_datasets.py summary

Validate what is present locally:

python scripts/prepare_datasets.py validate --write-manifest data/external/manifest.json

Convert one KITTI Velodyne frame to PLY:

python scripts/prepare_datasets.py convert-kitti-bin \
  --input data/external/kitti/velodyne/000000.bin \
  --output data/external/kitti/velodyne/000000.ply

Sample a ModelNet OFF mesh to an XYZ point cloud:

python scripts/prepare_datasets.py convert-modelnet-off \
  --input data/external/modelnet_small/chair.off \
  --output data/external/modelnet_small/sample.xyz \
  --points 2048

Create a deterministic Bunny source/target pair from one PLY:

python scripts/prepare_datasets.py make-bunny-pair \
  --input data/external/stanford/bunny/bunny.ply \
  --output-dir data/external/stanford/bunny_pair

Write checksums for reproducibility:

python scripts/prepare_datasets.py checksum \
  --input data/external \
  --output data/external/manifest.json

Real Data Examples

python examples/real_bunny_registration.py \
  --data-dir data/external/stanford/bunny_pair \
  --output-dir outputs/real_bunny

python examples/kitti_lidar_segmentation.py \
  --frame data/external/kitti/velodyne/000000.bin \
  --output-dir outputs/kitti_segmentation

python examples/modelnet_primitive_demo.py \
  --input data/external/modelnet_small/sample.xyz \
  --output-dir outputs/modelnet_demo

Each example exits with code 2 and a specific preparation hint if the expected data is missing.

Verification

Use checksums instead of committing large files:

python scripts/prepare_datasets.py checksum \
  --input data/external \
  --output data/external/manifest.json

The manifest records relative path, byte size, and SHA256 digest. This makes it possible to reproduce benchmark inputs while keeping the repository small.