HR-Depth: High Resolution Self-Supervised Monocular Depth Estimation

Self-supervised learning shows great potential in monocular depth estimation, using image sequences as the only source of supervision. Although people try to use high-resolution for accuracy prediction has not been significantly improved. In this work, we find core reason comes from inaccurate estimation large gradient regions, making bilinear interpolation error gradually disappear resolution increases. To obtain more accurate it is necessary features with spatial and semantic information. Therefore, present an improved DepthNet, HR-Depth, two effective strategies: (1) re-design skip-connection DepthNet get better (2) propose feature fusion Squeeze-and-Excitation(fSE) module fuse efficiently. Using Resnet-18 encoder, HR-Depth surpasses all previous state-of-the-art(SoTA) methods least parameters at both high low resolution. Moreover, SoTA are based on fairly complex deep networks a mass which limits their real applications. Thus also construct lightweight network uses MobileNetV3 encoder. Experiments show that can perform par many models like Monodepth2 only20%parameters. All codes will be available https://github.com/shawLyu/HR-Depth.