Skip to content

aimi-lab/RAFT-3D

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

RAFT-3D: Scene Flow using Rigid-Motion Embeddings

RAFT-3D: Scene Flow using Rigid-Motion Embeddings
Zachary Teed and Jia Deng, CVPR 2021

@inproceedings{teed2021raft3d,
  title={RAFT-3D: Scene Flow using Rigid-Motion Embeddings},
  author={Teed, Zachary and Deng, Jia},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2021},
}

Installation

RAFT-3D requires our LieTorch package to be installed. Please see https://github.com/princeton-vl/lietorch for instructions (requires PyTorch >= 1.6). All Cuda kernels required for RAFT-3D will automatically be installed with LieTorch.

Additional Requirements:

  • OpenCV
  • matplotlib
  • tqdm
  • Scikit-Sparse (only required for raft3d_laplacian)

Usage

Pretrained Models:

Demo:

You can run the demo to predict a SE3 field from a pair of RGB-D frames

python scripts/demo.py --network=raft3d.raft3d --model=raft3d.pth

or using the network with bi-Laplacian smoothing

python scripts/demo.py --network=raft3d.raft3d_bilaplacian --model=raft3d_laplacian.pth

FlyingThings3D Evaluation:

Evaluation requires the FlyingThings3D dataset to be downloaded. You also need to download the preprocessed FlowNet3D testing data from things_test_data.pickle.

Our code expects the following directory structure:

├── datasets
    ├── things_test_data.pickle
    ├── FlyingThings3D
        ├── frames_cleanpass
        ├── frames_finalpass
        ├── disparity
        ├── disparity_change
        ├── optical_flow
        ├── camera_data

You can then evaluate one of the pretrained models

python scripts/evaluation.py --network=raft3d.raft3d --model=raft3d.pth
python scripts/evaluation.py --network=raft3d.raft3d_bilaplacian --model=raft3d_laplacian.pth

KITTI Evaluation:

First download the KITTI Scene Flow dataset. You also need to download disparity estimates from GA-Net disp_ganet_training and disp_ganet_testing. Place the GA-Net estimates in the respective training and testing folders such that the directory structure is as follows

├── datasets
    ├── KITTI
        ├── devkit
        ├── training
            ├── disp_ganet_training
            ...
        ├── testing
            ├── disp_ganet_testing
            ...

You can then run the KITTI submission script

python scripts/kitti_submission.py --network=raft3d.raft3d_bilaplacian --model=raft3d_kitti.pth

Training

Training requires a GPU with 24Gb of memory. First download the required datasets for training and evaluation.

Training on FlyingThings3D

python -u scripts/train_things.py --name=r3d --batch_size=4 --lr=0.0002 --network=raft3d.raft3d_bilaplacian

Finetuning on KITTI

python -u scripts/train_kitti.py --name=r3d_kitti --batch_size=4 --lr=0.0001 --network=raft3d.raft3d_bilaplacian --ckpt=checkpoints/r3d_200000.pth

You can create a submission to the KITTI benchmark by running

python scripts/kitti_submission.py --model=checkpoints/r3d_kitti_050000.pth --network=raft3d.raft3d_bilaplacian

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%