Dual-branch Monocular Depth Estimation Method with Attention Mechanism

In order to solve the problem of fuzzy edge of near objects and fuzzy and inaccurate depth of local objects, a two-branch monocular depth estimation method with fused attention mechanism is proposed in this paper. In this method, the idea of classification regression is adopted. Transformer branch is used to fully extract the global remote context relationship, and CNN branch is used to process the dense texture area to extract local features. A feature fusion module based on cross-attention is designed to fully integrate the features of both branches and self-adaptively divide the depth interval to obtain the final depth map. At the same time, in order to make up for the inherent shortcomings of the two-branch structure baseline network, a dense void convolution pyramid module is proposed to further extract multi-scale features, and a parallel channel and location attention module is proposed to establish channel correlation and location correlation. Two public datasets, KITTI and NYU-V2, were tested, and the accuracy of the lowest threshold values reached 97.6% and 92.7%, respectively. The qualitative and quantitative results show that this method can effectively solve the problem of local object depth ambiguity and obtain high-quality scene depth map.