Codes for ECCV2022 paper "What matters in supervised 3D scene flow"

Codes for ECCV2022 paper "What matters in supervised 3D scene flow"

What Matters for 3D Scene Flow Network (ECCV2022)

This is the official implementation of our ECCV 2022 paper:

"What Matters for 3D Scene Flow Network"

Guangming Wang, Yunzhe Hu, Zhe Liu, Yiyang Zhou, Masayoshi Tomizuka, Wei Zhan, and Hesheng Wang



  • Python 3.6.9
  • PyTorch 1.8.0
  • CUDA 11.1
  • numba
  • tqdm

Data preprocessing

Data without occlusion

For fair comparison with previous methods, we adopt the preprocessing steps in HPLFlowNet. Please refer to repo. We also copy the preprocessing instructions here for your reference.

  • 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:
python3 data_preprocess/ --raw_data_path RAW_DATA_PATH --save_path SAVE_PATH/FlyingThings3D_subset_processed_35m --only_save_near_pts
python3 data_preprocess/ RAW_DATA_PATH SAVE_PATH/KITTI_processed_occ_final

Data with occlusion

The data with occlusion is prepared by FlowNet3D. Please refer to their repo. You can directly download their prepared data from here and here, then unzip to SAVE_PATH.


Install pointnet2 library

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

cd pointnet2
python install
cd ../

Train on non-occluded data

Set data_root in config_train.yaml to SAVE_PATH in the data preprocess section. Then run

python config_train.yaml

After training the model with a quarter dataset, you can finetune the model with the full dataset and achieve a better results by running the following command. Remember to set pretrain in config_train_finetune.yaml as the path to the pretrained weights.

python config_train_finetune.yaml

Evaluate on non-occluded data

We provide pretrained weights in pretrain_weights (EPE3D: 0.0282).

Set data_root and in config_evaluate.yaml to SAVE_PATH in the data preprocess section, and specify dataset in the script . Then run

python config_evaluate.yaml

Train on occluded data

Similarly, specify data_root in config_train_occlusion.yaml. Then run

python config_train_occlusion.yaml

Evaluate on occluded data

We provide pretrained weights in pretrain_weights (EPE3D: 0.0636).

Specify data_root and dataset in config_evaluate_occlusion.yaml. Then run

python config_evaluate_occlusion.yaml

Quantitative results

On non-occluded data results_non_occluded

On occluded data results_occluded


If you find our work useful in your research, please cite:

  title={What Matters for 3D Scene Flow Network},
  author={Wang, Guangming and Hu, Yunzhe and Liu, Zhe and Zhou, Yiyang and Tomizuka, Masayoshi and Zhan, Wei and Wang, Hesheng},
  booktitle={European Conference on Computer Vision},


We thank the following open-source projects for the help of the implementations.

Download Source Code

Download ZIP
Aug 18, 2022