Asymmetric convolution kernel for deep optical flow estimation

Deep optical flow estimation is a hot topic in computer vision in recent years. In this paper, an asymmetric convolution kernel is introduced to improve the accuracy of optical flow. Then, in order to alleviate the rasterization problem caused by dilated convolution in optical flow results, a method of hybrid dilated convolution is introduced. At the same time, the proposed method can keep a small number of parameters without using additional occlusion and bidirectional information. Finally, experiments on the open standard datasets MPI-Sintel and KITTI-15 are carried out, and the results demonstrate the effectiveness of the proposed approach.