PointDSC

To set up the PointDSC environment, use the provided environment.yml file with conda.

conda env create -f environment.yml
conda activate pointdsc

Training

Run the training script:

python train_KITTI.py

Testing

Run the evaluation script. Replace [exp_id] with your snapshot folder name:

python evaluation/test_KITTI.py --chosen_snapshot [exp_id] --use_icp False

To test traditional outlier rejection baselines, use baseline_scripts/baseline_KITTI.py.

To run multiway registration experiments, you must first download the Augmented_ICL dataset from the official website. The files must be organized into a specific directory structure where each scene (e.g., livingroom1-simulated) contains color and depth subdirectories containing .png files, accompanied by a trajectory file (e.g., livingroom1-traj.txt) in the parent directory.

├── livingroom1-simulated
│   ├── color
│   │   └── *.png
│   └── depth
│       └── *.png
├── livingroom1-traj.txt
├── livingroom2-simulated
│   ├── color
│   └── depth
├── livingroom2-traj.txt
├── office1-simulated
│   ├── color
│   └── depth
├── office1-traj.txt
├── office2-simulated
│   ├── color
│   └── depth
└── office2-traj.txt

Training

Run the training script:

python train_3dmatch.py

Training configurations can be modified in config.py.

Testing

Run the evaluation script. Replace [exp_id] with your snapshot folder name (e.g., PointDSC_3DMatch_release):

python evaluation/test_3DMatch.py --chosen_snapshot [exp_id] --use_icp False

Results are saved in logs/. To test traditional outlier rejection baselines, use baseline_scripts/baseline_3DMatch.py.

If you intend to use FCGF as the 3D local descriptor, you must perform additional installation steps:

1. Install MinkowskiEngine v0.5.0.
2. Download the FCGF model (pretrained on 3DMatch) from this link.

To use the 3DMatch dataset, ensure your data folder follows this structure:

.                          
├── fragments                 
│   ├── 7-scene-redkitechen/       
│   ├── sun3d-home_at-home_at_scan1_2013_jan_1/      
│   └── ...                
├── gt_result                   
│   ├── 7-scene-redkitechen-evaluation/   
│   ├── sun3d-home_at-home_at_scan1_2013_jan_1-evaluation/ 
│   └── ...                   
├── threedmatch            
│   ├── *.npz
│   └── *.txt                            

To reduce training time, pre-compute 3D local descriptors (FCGF or FPFH) using the provided scripts so that input correspondences can be built via NN search during training:

• For FCGF: misc/cal_fcgf.py
• For FPFH: misc/cal_fpfh.py

Execute the multiway registration test using the test_multi_ate.py script. You must specify a --chosen_snapshot to indicate which model snapshot to use. Upon completion, the script saves and visualizes the combined point cloud.

python multiway/test_multi_ate.py --chosen_snapshot [snapshot]

Local descriptors are computed for the generated .ply files to create putative correspondences for the network.

• To compute FPFH descriptors, run python misc/cal_fpfh.py.
• To compute FCGF descriptors, run python misc/cal_fcgf.py.

Note: If using FCGF, you must first download the pre-trained weights from this link.

python misc/cal_fpfh.py
# python misc/cal_fcgf.py

Use the make_fragments.py script to convert RGB-D sequences into point cloud fragments. You must provide a JSON configuration file for the specific scene. The resulting .ply files will be stored in a fragments/ directory within the scene's path.

python multiway/make_fragments.py redwood_simulated/livingroom1-simulated.json

To evaluate PointDSC on the 3DLoMatch benchmark, use the following command. If using the predator descriptor, you must first follow the official OverlapPredator instructions to extract features.

python evaluation/test_3DLoMatch.py --chosen_snapshot [exp_id] --descriptor [fcgf/predator] --num_points 5000

You can run a demo to extract dense FPFH descriptors for two point clouds and register them using PointDSC. The demo uses files in the demo_data folder.

Note on Models:

• Use the model pretrained on 3DMatch for indoor RGB-D scans.
• Use the model pretrained on KITTI for outdoor LiDAR scans.

Note on Descriptors:

• For challenging cases, it is recommended to use learned feature descriptors like FCGF or D3Feat instead of FPFH.

python demo_registration.py --chosen_snapshot [PointDSC_3DMatch_release/PointDSC_KITTI_release] --descriptor [fcgf/fpfh]