zju3dv / hybrid3d

Hybrid3D: Learning 3D Hybrid Features with Point Clouds and Multi-View Images for Point Cloud Registration

Installation

conda env create -f environment.yml
conda activate hf

Data Preparation

We provide scene fragment fusion in fuse_scene_fragment.py, fragment indices generation in generate_fragment_indices.py and the pre-computation of overlapping area in generate_overlapping_areas.py.

The preprocessed data and the pre-trained model would be avaiable in the future.

Training

# stage 1: training 2d coordinates
python train.py -c config/server_3dmatch_coord.yaml -i rgbd_stage
# stage 2: training 3d fusion
python train.py -c config/server_3dmatch_fusion.yaml --start_epoch 1 -r saved/model_xxx/checkpoint.pth  -i fusion_stage

Testing

See run_registration.sh and eval_registration_recall.sh in evaluation folder.