Optical flow estimation combining with objects edge features

Optical flow estimation has been concerned with biological plausibility since the optical flow was drawn from psychology. In this Letter, optical flow estimation combining with objects edge features is proposed. Inspired of functional division and collaboration of edge detection in primary visual cortex and motion information acquisition in a motion-selective area in the human visual cortex, a neural sub-network is introduced to extract motion edge information as primary visual cortex does, and the other neural sub-network is utilised to perceive motion information as a motion-selective area in visual cortex does. The inference sub-network in the authors’ model used augmentation features composed of objects, edge features and general features as inputs to predict the optical flow. In experiments on large displacements benchmark datasets, their approach showed promising results to improve the performance of optical flow comparing with the other models based on the visuomotor perception of visual cortex processing mechanism.