Analyzing motion between two consecutive images is one of the fundamental tasks in computer vision. In the lack of labeled data, the loss functions are split into consistency and smoothness, allowing for self-supervised training. This paper focuses on the cost function derivation and presents an unrolling iterative approach, transferring the hard L1 smoothness constraint into a softer multi-layer iterative scheme. More accurate gradients, especially near non-differential positions, improve the network's convergence, providing superior results on tested scenarios. We report state-of-the-art results on both MPI Sintel and KITTI 2015 unsupervised optical flow benchmarks. The provided approach can be used to enhance various architectures and not limited just to the presented pipeline.

Unsupervised Optical Flow Using Cost Function Unrolling