FlowNet2 (TensorFlow)

This repo includes FlowNetC, S, CS, CSS, CSS-ft-sd, SD, and 2 for TensorFlow. Most part are from this repo, and we have made some modifications:

• It can deal with arbitrary size of input now.
• After installation, just copy the whole folder FlowNet2_src to your codebase to use. See demo.py for details.

Environment

This code has been tested with Python3.6 and TensorFlow1.2.0, with a Tesla K80 GPU. The system is Ubuntu 14.04.

Installation

You must have CUDA installed: make all

Note: you might need to modify this line, according to the GPU you use.

Download weights

To download the weights for all models (4.4GB), run the download.sh script in the FlowNet2_src/checkpoints directory. All test scripts rely on these checkpoints to work properly.

Inference mode

python demo.py

If installation is successful, you should see the following:

Notice that the model itself will handle the RGB to BGR operation for you. And please be care about your input scale and datatype.

Performance (w/o fine-tuning)

Model	KITTI2012 Train EPE	KITTI2015 Train EPE	KITTI2015 Train F1	Sintel Final Train EPE
FlowNetS	7.2457	14.0753	0.5096	3.9140
FlowNetC	5.9793	11.8957	0.4509	3.1001
FlowNet2	4.3167	10.9869	0.3241	2.1592

FlowNetS and FlowNetC are better than paper, but FlowNet2 is slightly worse.

TODO

• Add fine-tune mode
• Remove the training_schedule variable from inference mode.

Reference

[1] E. Ilg, N. Mayer, T. Saikia, M. Keuper, A. Dosovitskiy, T. Brox FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks, IEEE Conference in Computer Vision and Pattern Recognition (CVPR), 2017.

Acknowledgments

As noted in the beginning, most part are from sampepose/flownet2-tf