liuguoyou/raft

RAFT

This repository contains the source code for our paper:

RAFT: Recurrent All Pairs Field Transforms for Optical Flow
ECCV 2020
Zachary Teed and Jia Deng

Requirements

The code has been tested with PyTorch 1.6 and Cuda 10.1.

conda create --name raft
conda activate raft
conda install pytorch=1.6.0 torchvision=0.7.0 cudatoolkit=10.1 -c pytorch
conda install matplotlib
conda install tensorboard
conda install scipy
conda install opencv

Demos

Pretrained models can be downloaded by running

./download_models.sh

or downloaded from google drive

You can demo a trained model on a sequence of frames

python demo.py --model=models/raft-things.pth --path=demo-frames

Required Data

To evaluate/train RAFT, you will need to download the required datasets.

• FlyingChairs
• FlyingThings3D
• Sintel
• KITTI
• HD1K (optional)

By default datasets.py will search for the datasets in these locations. You can create symbolic links to wherever the datasets were downloaded in the datasets folder

├── datasets
    ├── Sintel
        ├── test
        ├── training
    ├── KITTI
        ├── testing
        ├── training
        ├── devkit
    ├── FlyingChairs_release
        ├── data
    ├── FlyingThings3D
        ├── frames_cleanpass
        ├── frames_finalpass
        ├── optical_flow

Evaluation

You can evaluate a trained model using evaluate.py

python evaluate.py --model=models/raft-things.pth --dataset=sintel --mixed_precision

Training

We used the following training schedule in our paper (2 GPUs). Training logs will be written to the runs which can be visualized using tensorboard

./train_standard.sh

If you have a RTX GPU, training can be accelerated using mixed precision. You can expect similiar results in this setting (1 GPU)

./train_mixed.sh

(Optional) Efficent Implementation

You can optionally use our alternate (efficent) implementation by compiling the provided cuda extension

cd alt_cuda_corr && python setup.py install && cd ..

and running demo.py and evaluate.py with the --alternate_corr flag Note, this implementation is somewhat slower than all-pairs, but uses significantly less GPU memory during the forward pass.