Learning Optical Flow via Deformable Convolution and Feature Pyramid Networks

Convolutional neural networks have powerful learning capabilities which can simulate various complex mapping relationships, so they are widely used in various fields of computer vision. Nowadays, learning optical flow based on CNN has achieved remarkable results. However, optical flow learning approaches still require further research and development. Firstly, the large displacement problem in the traditional methods still exists in optical flow learning approaches. What's more, the development of optical flow learning approaches are limited to a certain extent because of convolutional neural networks lacks the ability to handle geometric transformations, which result from their fixed geometry structure. In order to solve above problems, we introduce the deformation convolution into the optical flow estimation network, which can adaptively adjust the shape of the receptive field to learn different levels and different types of motion information. In addition, the feature pyramid is also integrated into the network to further enhance the learning ability of optical flow learning network. The experimental results surpass many state-of-the-art algorithms in the sintel dataset, which proves the effectiveness of our approach.