MMAFlow: Matching-Guided Motion Aggregation for Optical Flow Estimation

Abstract

In recent years, deep learning has achieved promising results in optical flow estimation. However, these learning-based methods encounter challenges when dealing with occlusions, especially in the two-view setting. To further improve performance in occluded regions, we introduce an optical flow framework named MMAFlow. Specifically, MMAFlow first estimates an occlusion mask and a coarse flow that provide priors of occlusions and large motions, and then passes motion information from non-occluded pixels to occluded pixels by applying occlusion-aware motion aggregation. Moreover, the aggregated motion information will be utilized by subsequent flow optimizers to obtain the final flow. The experimental results on challenging Sintel and KITTI benchmarks verified the superiority of our proposed MMAFlow, which gains a notable reduction of 10.8% in the average end-point error on Sintel final pass compared to the baseline model, and achieves an F1-all error rate of 4.49% on KITTI 2015.