Awesome

RigidFlow

This is the PyTorch code for RigidFlow: Self-Supervised Scene Flow Learning on Point Clouds by Local Rigidity Prior (CVPR2022). The code is created by Ruibo Li (ruibo001@e.ntu.edu.sg).

Updates

We release an extended version of RigidFlow on Self-Supervised 3D Scene Flow Estimation and Motion Prediction using Local Rigidity Prior (T-PAMI 2024). You can also check out the arXiv version at RigidFlowPP-arXiv.

The code for self-supervised scene flow estimation can be found in RigidFlowPP, while the code for self-supervised motion prediction is available in RigidFlowPP-Motion.

Prerequisities

• Python 3.6.13
• NVIDIA GPU + CUDA CuDNN
• PyTorch (torch == 1.4.0)
• tqdm
• sklearn
• pptk
• yaml

Create a conda environment for RigidFlow:

conda create -n RigidFlow python=3.6.13
conda activate RigidFlow
conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.1 -c pytorch
pip install tqdm pptk PyYAML

Compile the furthest point sampling, grouping and gathering operation for PyTorch. We use the operation from this repo.

cd lib
python setup.py install
cd ../

Install & complie supervoxel segmentation method:

cd Supervoxel_utils
g++ -std=c++11 -fPIC -shared -o main.so main.cc
cd ../

More details about the supervoxel segmentation method, please refer to Supervoxel-for-3D-point-clouds.

Data preprocess

By default, the datasets are stored in SAVE_PATH.

FlyingThings3D

Download and unzip the "Disparity", "Disparity Occlusions", "Disparity change", "Optical flow", "Flow Occlusions" for DispNet/FlowNet2.0 dataset subsets from the FlyingThings3D website (we used the paths from this file, now they added torrent downloads) . They will be upzipped into the same directory, RAW_DATA_PATH. Then run the following script for 3D reconstruction:

python data_preprocess/process_flyingthings3d_subset.py --raw_data_path RAW_DATA_PATH --save_path SAVE_PATH/FlyingThings3D_subset_processed_35m --only_save_near_pts

This dataset is denoted FT3D_s in our paper.

KITTI

• KITTI scene flow data provided by HPLFlowNet

Download and unzip KITTI Scene Flow Evaluation 2015 to directory RAW_DATA_PATH. Run the following script for 3D reconstruction:

python data_preprocess/process_kitti.py RAW_DATA_PATH SAVE_PATH/KITTI_processed_occ_final

This dataset is denoted KITTI_s in our paper.

• KITTI scene flow data provided by FlowNet3D

Download and unzip data processed by FlowNet3D to directory SAVE_PATH. This dataset is denoted KITTI_o in our paper.

• Unlabeled KITTI raw data

In our paper, we use raw data from KITTI for self-supervised scene flow learning. We release the unlabeled training data here for download. This dataset is denoted KITTI_r in our paper.

Evaluation

Set data_root in each configuration file to SAVE_PATH in the data preprocess section.

Trained models

Our trained models can be downloaded from model trained on FT3D_s and model trained on KITTI_r.

Testing

• Model trained on FT3D_s

When evaluating this pre-trained model on FT3D_s testing data, set dataset to FT3D_s_test. And when evaluating this pre-trained model on KITTI_s data, set dataset to KITTI_s_test. Then run:

python evaluate.py config_evaluate_FT3D_s.yaml

• Model trained on KITTI_r

Evaluate this pre-trained model on KITTI_o:

python evaluate.py config_evaluate_KITTI_o.yaml

Training

Set data_root in each configuration file to SAVE_PATH in the data preprocess section.

• Train model on FT3D_s with 8192 points as input:

python train_FT3D_s.py config_train_FT3D_s.yaml

• Train model on KITTI_r with 2048 points as input:

python train_KITTI_r.py train_KITTI_r.yaml

Citation

If you find this code useful, please cite our paper:

@inproceedings{li2022rigidflow,
  title={RigidFlow: Self-Supervised Scene Flow Learning on Point Clouds by Local Rigidity Prior},
  author={Li, Ruibo and Zhang, Chi and Lin, Guosheng and Wang, Zhe and Shen, Chunhua},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={16959--16968},
  year={2022}
}

@article{li2024self,
  title={Self-Supervised 3D Scene Flow Estimation and Motion Prediction using Local Rigidity Prior},
  author={Li, Ruibo and Zhang, Chi and Wang, Zhe and Shen, Chunhua and Lin, Guosheng},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2024},
  publisher={IEEE}
}

Acknowledgement

Our code is based on Self-Point-Flow, HPLFlowNet, FLOT, and PointPWC. Our implemented FLOT model is based on FLOT and flownet3d_pytorch. The supervoxel segmentation method is based on Supervoxel-for-3D-point-clouds.