gonglixue / RAFT-tf

ECCV2020-RAFT(Recurrent All Pairs Field Transforms for Optical Flow) tensorflow/tensorpack implementation

RAFT-tensorflow/tensorpack

This repository contains the tensorflow-implementation(with tensorpack) for this paper:

RAFT:Recurrent All Pairs Field Transforms for Optifal Flow

Official Pytorch Implementation

TODO

• Basic inference code.
• grid_sample align_corners=True in tf-implementation.
• Add cuda extension for efficent correlation calculation.
• Check if the batch_size could be free.
• Reproduce the training process.

Requirements

tensorflow-gpu >= 1.14
opencv-python
numpy
tensorpack
CUDA 10.1

Demo

1. Download the pretrained model from GoogleDrive to release_weight folder. The *.npz files are converted from the official pytorch *.pth model provided in the official repository.

2. Run the inference demo:

bash ./infer_image.sh

or

python infer_raft.py --im1 frame_0016.png --im2 frame_0017.png --load release_weight/raft-things.npz

python infer_raft.py --im1 frame_0016.png --im2 frame_0017.png --load release_weight/raft-small.npz --small

Note

The inference result of my tensorflow implementation is as below. There is still a few of differences from the official implementation. I will continue to follow up.

My tf-implementation	Official pytorch-implementation