GenFlow3D

[ICCV 2025] Code for "GenFlow3D: Generative scene flow estimation and prediction on point cloud sequences". Hanlin Li, Wenming Weng, Yueyi Zhang†, Zhiwei Xiong

Preproration

To run our model, install the PointNet2 library first as follows.

cd pointnet2
python setup.py install
cd ..

After this, we need to prepare the sequential scene flow datasets nuScenes and Argoverse 2.

nuScenes

The code to preprocess the nuScenes dataset is in ./data/nuscenes. We follow SLIM to generate the scene flow ground truth.

Download the full nuScenes v1.0 trainval dataset from the official website and unzip them. The directory structure should be as follows.

your_path_to_nuscenes--|
                       |--maps
                       |--samples
                       |--sweeps
                       |--v1.0-trainval
                       |--LICENSE

Then, run the code file ./data/nuscenes/create_nuscenes_seq_length.py:

cd ./data/nuscenes
python create_nuscenes_seq_length.py --nusc_root your_path_to_nuscenes --path_out OUTPUT_PATH_NUSCENES
cd ../..

The processed nuScenes dataset is in OUTPUT_PATH_NUSCENES.

Argoverse 2

The code to preprocess the Argoverse 2 dataset is in ./data/av2. We use the official method from av2-api to generate the scene flow ground truth.

Download the training parts and validation parts of the Argoverse 2 Sensor Dataset from the official website and unzip them. The directory structure should be as follows.

your_path_to_av2--|
                  |--av2--|
                          |--sensor--|
                                     |--train--|
                                     |         |--00a6ffc1-6ce9-3bc3-a060-6006e9893a1a
                                     |         |--... 
                                     |--val--|
                                     |       |--0aa4e8f5-2f9a-39a1-8f80-c2fdde4405a2
                                     |       |--...

Then run the code file ./data/av2/create_av2_seq.py：

cd ./data/av2
python create_av2_seq.py --data_root your_path_to_av2 --out_root OUTPUT_PATH_AV2 --split_name train
python create_av2_seq.py --data_root your_path_to_av2 --out_root OUTPUT_PATH_AV2 --split_name val
cd ../..

The processed nuScenes dataset is in OUTPUT_PATH_AV2.

Evaluation

The pretrianed models are in ./pretrain.

To evaluate on nuScenes, first, set the data_root term in the config file test_cfg.yaml as your path to the processed nuScenes dataset. Then run the test code test.py:

python test.py

To evaluate on Argoverse 2, similarly, set the data_root term in the config file test_cfg_occ.yaml as your path to the processed Argoverse 2 dataset. Then run the test code test_occ.py:

python test_occ.py

Training

To train on nuScenes, set the data_root term in the config file train_cfg.yaml as your path to the processed nuScenes dataset. Then run the training code train.py:

python train.py

To train on Argoverse 2, set the data_root term in the config file train_cfg_occ.yaml as your path to the processed Argoverse 2 dataset. Then run the training code train_occ.py:

python train_occ.py

Acknowledgement

Our code is based on flownet3d_pytorch, PointPWC-Net, Bi-PointFlowNet, MSBRN and DifFlow3D.

The code to preprocess the nuScenes dataset is based on SLIM and the code to preprocess the Argoverse 2 dataset is based on av2-api.

The PointNet2 library is from this repository.