Self-supervised Depth Estimation with High Resolution Features and Non-local Information

Depth estimation is one of the most challenging tasks in computer vision, especially in self-supervised learning ways without restrictions of high-cost labels. Self-supervised depth estimation aims to infer three-dimensional space structures from two-dimensional planar images, only taking image pairs or sequences as supervision. Most existing methods adopt the encoder-decoder framework with skip-connection and recover the high-resolution depth maps from high-resolution low-level and low-resolution high-level feature maps. However, it is proved that high-resolution high-level feature maps, which are sensitive to illumination, color, texture, etc., are necessary for depth estimation. In this paper, we present a novel approach to extract high-level feature maps at all scales and introduce a self-attention mechanism to consider non-local features. The main improvements of our proposed method are two-fold:1) we combined the high-resolution feature extraction sub-network and extract high-resolution high-level features by connecting the high-to-low resolution convolution streams in parallel; 2) we embed the self-attention module with the features pyramid module(FPA) to obtain general context at large-scale features. The experiments evaluated on the KITTI benchmark have demonstrated that our network outperforms most existing methods and produces more accurate depth maps.