Monocular Scene Flow Estimation with Convolution Vision Transformer

The process of artificial intelligence processing is inseparable from the perception of information in the scene, and the monocular scene flow estimation combining the optical flow and depth can reflect the motion information in the scene. In this paper, a monocular scene flow estimation network with convolution vision transformer(CVT) is proposed. Specifically, introducing CVT into the feature extraction part of the monocular scene flow estimation network can effectively capture local information while dynamically paying attention to global context, which solves the problems of large local redundancy and complex global dependencies in computer vision data. In addition, a single rectified linear unit(RELU) activation is used in the CVT to provide pixel-wise modeling capabilities to enrich the detail feature. On the KITTI driving dataset, the experimental results show that the proposed method improves the performance of the estimation of the network and has the competitive advantage.