Dual-Attention Mechanism for Monocular Depth Estimation

Depth estimation stands as a fundamental task in computer vision, playing a crucial role in accurately assessing obstacle distances. This paper presents a supervised learning-based monocular depth estimation network using dual-attention mechanism. First, a spatial self-attention mechanism is used to extract disparity features from continuous images, where disparity features includes shape, color, texture, and edge features of multiscale objects. Subsequently, the sequential convolutional attention mechanism is used to predict depth, employing multi-layer stacking to capture richer detailed information and enhance depth estimation capabilities. Finally, the developed monocular depth estimation method was validated and compared using the KITTI Eigen split public database for quantitative results in this paper.