CNN-based Patch Matching for Optical Flow with Thresholded Hinge Embedding Loss

TL;DR

This paper introduces a CNN-based patch matching method with a novel thresholded hinge embedding loss for optical flow, achieving state-of-the-art results on multiple benchmarks and improving training efficiency.

Contribution

It presents a new thresholded loss function for Siamese networks and a novel multi-scale CNN feature calculation method, enhancing optical flow estimation.

Findings

The thresholded hinge embedding loss outperforms existing losses.

Training speed is doubled with the new loss.

Achieved state-of-the-art results on KITTI and MPI-Sintel datasets.

Abstract

Learning based approaches have not yet achieved their full potential in optical flow estimation, where their performance still trails heuristic approaches. In this paper, we present a CNN based patch matching approach for optical flow estimation. An important contribution of our approach is a novel thresholded loss for Siamese networks. We demonstrate that our loss performs clearly better than existing losses. It also allows to speed up training by a factor of 2 in our tests. Furthermore, we present a novel way for calculating CNN based features for different image scales, which performs better than existing methods. We also discuss new ways of evaluating the robustness of trained features for the application of patch matching for optical flow. An interesting discovery in our paper is that low-pass filtering of feature maps can increase the robustness of features created by CNNs. We proved the competitive performance of our approach by submitting it to the KITTI 2012, KITTI 2015 and MPI-Sintel evaluation portals where we obtained state-of-the-art results on all three datasets.