GeDi: Learning general and distinctive 3D local deep descriptors for point cloud registration - IEEE T-PAMI

Fork of GeDi. Paper (pdf)

Currently built for:

• Windows/Linux OS
• Python: 3.9/3.10/3.11/3.12
• CPU
• CUDA: 11.8/12.1/12.4/12.6/12.8/12.9
• CUDA_ARCH: 6.0/6.1/7.5/8.6/8.9

Wheel build for developers

Required: Install pixi.

To build the wheel of GeDi, you first need to choose the Python/CPU/CUDA versions with which the wheel will be built (example: "py312-cu124"). Replace PIXI_ENV placeholder in the commands below with your environment name.

Run on Windows:

pixi run -e PIXI_ENV wheel-build-windows

Run on Linux:

pixi run -e PIXI_ENV wheel-build-linux

The wheel will be created in ./Release folder.

Code overview: registration.py

The registration function compute_registration_matrix is structured as follows:

def compute_registration_matrix(
    pcd0_points: np.ndarray,
    pcd1_points: np.ndarray,
    samples_per_batch: int = 500,
    samples_per_patch_lrf: int = 4000,
    samples_per_patch_out: int = 512,
    lrf_radius: float = 0.5,
    voxel_size: float = 0.01,
    patches_per_pair: int = 5000,
    max_correspondence_distance: float = 0.02,
    edge_length_checker: float = 0.9,
    distance_checker: float = 0.02,
    ransac_iterations: int = 4000000,
    ransac_correspondences_threshold: int = 40,
    visualize: bool = False,
    device: str = 'cuda'
) -> np.ndarray:

Inputs

• pcd0_points is a NumPy array of shape [N, 3] containing the N-points of the first point cloud;
• pcd1_points is a NumPy array of shape [M, 3] containing the M-points of the second point cloud;
• samples_per_batch is the number of samples taken from patches_per_pair that are processed in a single batch; the number of batches will therefore be patches_per_pair/samples_per_batch;
• samples_per_patch_lrf is the number of points to process with LRF;
• samples_per_patch_out is the number of points to sample for pointnet++;
• lrf_radius is the radius used to compute the Local Reference Frame (LRF) for descriptor computation;
• voxel_size is the voxel size used to downsample the point clouds before registration;
• patches_per_pair is the number of point patches randomly sampled per point cloud to compute descriptors;
• max_correspondence_distance is the maximum distance threshold for matching points during RANSAC;
• edge_length_checker is the edge length similarity threshold for the RANSAC correspondence checker;
• distance_checker is the distance similarity threshold for the RANSAC correspondence checker;
• ransac_iterations is the maximum number of RANSAC iterations;
• ransac_correspondences_threshold is the minimum number of correspondences required to consider RANSAC successful. After 3 unsuccessful attempts this threshold is lowered by 10 each cycle until Ransac becomes successful;
• visualize (bool) whether to visualize the clouds through Open3D visualizer;
• device (str) whether to use CPU (arg: 'cpu') or GPU (arg: 'cuda') to run registration. Note: Inference on CPU takes roughly 6× more time compared to GPU.

Output

• A np.ndarray 4×4 transformation matrix that aligns pcd0 to pcd1.

The function compute_registration_matrix_from_path provides the same registration logic as compute_registration_matrix, but loads the point clouds from paths instead of receiving them as NumPy arrays:

Differences in Inputs

• pcd0_path is the file path (string) to the first point cloud;
• pcd1_path is the file path (string) to the second point cloud;

All the other Inputs and the Output remain the same.

Use of registration.py from command line

If you want to test or debug locally the registration you can run (replace PIXI_ENV placeholder with your environment tag):

pixi run -e PIXI_ENV python registration.py --pcd0 /path/to/cloud0 --pcd1 /path/to/cloud1 --visualize --device cuda

Citations

@inproceedings{Poiesi2021,
  title = {Learning general and distinctive 3D local deep descriptors for point cloud registration},
  author = {Poiesi, Fabio and Boscaini, Davide},
  booktitle = {IEEE Trans. on Pattern Analysis and Machine Intelligence},
  year = {(early access) 2022}
}