The Optical Flow Estimation Method Based on the Attention Feature Enhancement Module

Optical flow estimation is one of the classic tasks in the field of vision. It calculates the motion of the target based on two continuous images. In most of the past work, the estimated optical flow often has errors in scenes of large motion or noise interference because the convolutional neural network structure often focuses on local features. To overcome the problems mentioned above, we propose a feature enhancement module that is based on the self-attention mechanism to enhance the dependencies of long-range features. Additionally, It can filter part of the noise, such as the interference of illumination in the input image. We evaluate our approach on standard benchmarks to verify motion estimation ability. To compare different results intuitively, we visualize flow fields for qualitative analysis. Eventually, we use the heat map to visualize the output of the attention layer to explore the mechanism of the attention algorithm.