UnFlowGAN: Unsupervised Learning of Optical Flow with Adversarial Learning

Abstract

Recently, unsupervised learning-based approaches for optical flow estimation have been actively researched. In unsupervised settings, the difference between the input image and the reconstructed image created from the estimated optical flow is minimized to learn the optical flow. Conventional learning methods mainly treated the difference as the pixel-by-pixel brightness error, which might lead to decreasing accuracy of the optical flow because the learning-strategy cannot take the textures into account sufficiently. To deal with the problem, in addition to the brightness error, we propose the introduction of adversarial learning into the evaluation of the input image and the reconstructed image. Our main contribution is that we develop a learning-strategy to capture the textures and the proposed method outperforms the conventional methods on the KITTI benchmarks.