Skip to content

Official repository for NeurIPS 2023 paper: GMSF: Global Matching Scene Flow

Notifications You must be signed in to change notification settings

ZhangYushan3/GMSF

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

GMSF: Global Matching Scene Flow


GMSF: Global Matching Scene Flow
Yushan Zhang, Johan Edstedt, Bastian Wandt, Per-Erik Forssén, Maria Magnusson, Michael Felsberg
NeurIPS 2023

Get started

Here are some demo results:

Figure_3.mp4

Installation:

Create a conda environment:

conda create -n GMSF python=3.8
conda activate GMSF
conda install pytorch torchvision pytorch-cuda=11.7 -c pytorch -c nvidia

Install other dependencies:

pip install opencv-python open3d tensorboard imageio numba

Dataset Preparation:

  1. FlyingThings3D(HPLFlowNet without occlusion / CamLiFlow with occlusion):

Download FlyingThings3D_subset. flyingthings3d_disparity.tar.bz2, flyingthings3d_disparity_change.tar.bz2, FlyingThings3D_subset_disparity_occlusions.tar.bz2, FlyingThings3D_subset_flow.tar.bz2, FlyingThings3D_subset_flow_occlusions.tar.bz2 and FlyingThings3D_subset_image_clean.tar.bz2 are needed. Then extract the files in /path/to/flyingthings3d such that the directory looks like

/path/to/flyingthings3d
├── train/
│   ├── disparity/
│   ├── disparity_change/
│   ├── disparity_occlusions/
│   ├── flow/
│   ├── flow_occlusions/
│   ├── image_clean/
├── val/
│   ├── disparity/
│   ├── disparity_change/
│   ├── disparity_occlusions/
│   ├── flow/
│   ├── flow_occlusions/
│   ├── image_clean/

Preprocess dataset using the following command:

cd utils
python preprocess_flyingthings3d_subset.py --input_dir /mnt/data/flyingthings3d_subset --output_dir flyingthings3d_subset
python preprocess_flyingthings3d_subset.py --input_dir /mnt/data/flyingthings3d_subset --output_dir flyingthings3d_subset_non-occluded --remove_occluded_points
  1. FlyingThings3D(FlowNet3D with occlusion):

The processed data is also provided here for download (total size ~11GB)

  1. KITTI(HPLFlowNet without occlusion):

First, download the following parts: Main data: data_scene_flow.zip Calibration files: data_scene_flow_calib.zip Unzip them and organize the directory as follows:

datasets/KITTI_stereo2015
├── testing
│   ├── calib_cam_to_cam
│   ├── calib_imu_to_velo
│   ├── calib_velo_to_cam
│   ├── image_2
│   ├── image_3
└── training
    ├── calib_cam_to_cam
    ├── calib_imu_to_velo
    ├── calib_velo_to_cam
    ├── disp_noc_0
    ├── disp_noc_1
    ├── disp_occ_0
    ├── disp_occ_1
    ├── flow_noc
    ├── flow_occ
    ├── image_2
    ├── image_3
    ├── obj_map

Preprocess dataset using the following command:

cd utils
python process_kitti.py datasets/KITTI_stereo2015/ SAVE_PATH/KITTI_processed_occ_final
  1. KITTI(FlowNet3D with occlusion):

The processed data is also provided here for download

  1. Waymo-Open(refer to FH-Net): Download the Waymo raw data from link_to_waymo_open_dataset,run the following command to extract point clouds, 3D annotations, poses and other information form raw data.
cd gmsf
python waymo_tools/waymo_extract.py

After extracting data, the folder structure is the same as below:

datasets
├── waymo-open
│   ├── scene_flow
│       ├── ImageSets
│       ├── train
│       ├── valid
├── train_extract
│   ├── 000
│   ├── 001
│   ├── ...
├── valid_extract
│   ├── 000
│   ├── 001
│   ├── ...

Then create scene flow data by:

python waymo_tools/create_data.py --dataset_type waymo

The datasets directory should be orginized as:

datasets
├── datasets_KITTI_flownet3d
│   ├── kitti_rm_ground
├── datasets_KITTI_hplflownet
│   ├── KITTI_processed_occ_final
├── FlyingThings3D_flownet3d
├── flyingthings3d_subset
│   ├── train
│   ├── val
├── flyingthings3d_subset_non-occluded
│   ├── train
│   ├── val
├── KITTI_stereo2015
│   ├── testing
│   ├── training
├── waymo-open
│   ├── scene_flow
│   ├── train_extract
│   ├── valid_extract

Traning and Testing:

Training (HPLFlowNet / CamLiFlow with occlusion):

cd gmsf
python main_gmsf.py \
    --checkpoint_dir checkpoints \
    --stage things_subset \
    --backbone DGCNN \
    --num_transformer_pt_layers 1 \
    --num_transformer_layers 10 \
    --feature_channels_point 128 \
    --lr 2e-4 --batch_size 8 --num_steps 600000

Training (HPLFlowNet without occlusion):

cd gmsf
python main_gmsf.py \
    --checkpoint_dir checkpoints \
    --stage things_subset_non-occluded \
    --backbone DGCNN \
    --num_transformer_pt_layers 1 \
    --num_transformer_layers 10 \
    --feature_channels_point 128 \
    --lr 2e-4 --batch_size 8 --num_steps 600000

Training (FlowNet3D with occlusion):

cd gmsf
python main_gmsf.py \
    --checkpoint_dir checkpoints \
    --stage things_flownet3d \
    --backbone DGCNN \
    --num_transformer_pt_layers 1 \
    --num_transformer_layers 10 \
    --feature_channels_point 128 \
    --lr 2e-4 --batch_size 8 --num_steps 600000

Training (Waymo-Open):

cd gmsf
python main_gmsf.py \
    --checkpoint_dir checkpoints \
    --stage waymo \
    --backbone DGCNN \
    --num_transformer_pt_layers 1 \
    --num_transformer_layers 10 \
    --feature_channels_point 128 \
    --lr 1e-4 --batch_size 8 --num_steps 600000

Testing (HPLFlowNet / CamLiFlow with occlusion):

cd gmsf
python main_gmsf.py --resume checkpoints/step_600000.pth \
    --stage things_subset \
    --backbone DGCNN \
    --num_transformer_pt_layers 1 \
    --num_transformer_layers 10 \
    --feature_channels_point 128 \
    --eval

Testing (HPLFlowNet without occlusion):

cd gmsf
python main_gmsf.py --resume checkpoints/step_600000.pth \
    --stage things_subset_non-occluded \
    --backbone DGCNN \
    --num_transformer_pt_layers 1 \
    --num_transformer_layers 10 \
    --feature_channels_point 128 \
    --eval

Testing (FlowNet3D with occlusion):

cd gmsf
python main_gmsf.py --resume checkpoints/step_600000.pth \
    --stage things_flownet3d \
    --backbone DGCNN \
    --num_transformer_pt_layers 1 \
    --num_transformer_layers 10 \
    --feature_channels_point 128 \
    --eval

Testing (Waymo-Open):

cd gmsf
python main_gmsf.py --resume checkpoints/step_600000.pth \
    --stage waymo \
    --backbone DGCNN \
    --num_transformer_pt_layers 1 \
    --num_transformer_layers 10 \
    --feature_channels_point 128 \
    --eval

Pretrained Checkpoints

Model trained on FTD_c: MODEL_FTDc

Model trained on FTD_o: MODEL_FTDo

Model trained on FTD_s: MODEL_FTDs

Model trained on Waymo: MODEL_Waymo

BibTeX

If you find our models useful, please consider citing our paper!

@article{zhang2023gmsf,
  title={GMSF: Global Matching Scene Flow},
  author={Zhang, Yushan and Edstedt, Johan and Wandt, Bastian and Forss{\'e}n, Per-Erik and Magnusson, Maria and Felsberg, Michael},
  journal={arXiv preprint arXiv:2305.17432},
  year={2023}
}

About

Official repository for NeurIPS 2023 paper: GMSF: Global Matching Scene Flow

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages