FlowNet: Learning Optical Flow with Convolutional Networks
Convolutional neural networks (CNNs) have recently been very successful in a variety of computer vision tasks, especially on those linked to recognition. Optical flow estimation has not been among the tasks CNNs succeeded at. In this paper we construct CNNs which are capable of solving the optical flow estimation problem as a supervised learning task. We propose and compare two architectures: a generic architecture and another one including a layer that correlates feature vectors at different image locations. Since existing ground truth data sets are not sufficiently large to train a CNN, we generate a large synthetic Flying Chairs dataset. We show that networks trained on this unrealistic data still generalize very well to existing datasets such as Sintel and KITTI, achieving competitive accuracy at frame rates of 5 to 10 fps.
| Models | Training datasets | FlyingChairs | Sintel (training) | KITTI2012 (training) | Log | Config | Download | |
| clean | final | EPE | ||||||
|---|---|---|---|---|---|---|---|---|
| FlowNetC | FlyingChairs | 1.78 | 3.60 | 4.93 | 7.55 | log | Config | Model |
| FlowNetC | Flying Chairs + FlyingThing3d subset | 2.57 | 2.74 | 4.52 | 5.42 | log | Config | Model |
| FlowNetC+ft | Flying Chairs + Sintel | 2.80 | 1.73 | 2.09 | 4.78 | log | Config | Model |
| FlowNetS | FlyingChairs | 2.03 | 4.25 | 5.64 | 8.81 | log | Config | Model |
| FlowNetS+ft | FlyingChairs + Sintel | 3.06 | 1.93 | 2.12 | 6.83 | log | Config | Model |
@inproceedings{dosovitskiy2015flownet,
title={Flownet: Learning optical flow with convolutional networks},
author={Dosovitskiy, Alexey and Fischer, Philipp and Ilg, Eddy and Hausser, Philip and Hazirbas, Caner and Golkov, Vladimir and Van Der Smagt, Patrick and Cremers, Daniel and Brox, Thomas},
booktitle={Proceedings of the IEEE international conference on computer vision},
pages={2758--2766},
year={2015}
}