Abstract:Current mainstream indoor self-supervised monocular depth estimation frameworks predominantly adopt either CNN (convolutional neural network) or ViT (vision Transformer) as a single backbone. CNN is constrained by limited receptive fields, which hinder the modeling of long-range dependencies and result in blurred depth estimation in weakly textured regions. Although ViT exhibits strong global modeling capability, edge and fine-grained details are easily lost at high-resolution feature stages, leading to object boundary distortions. To address these limitations, a parallel dual-branch network, termed Swin-DPEC, is proposed, which integrates the global context modeling capability of Swin-Transformer with the local detail extraction ability of a dual-path enhanced convolution (DPEC) module to achieve complementary advantages. The global branch of the network is constructed based on Swin-Transformer and exploits a hierarchical window attention mechanism to capture multi-scale long-range context, thereby providing geometric priors for low-texture regions. The local branch enhances the perception of high-frequency structures, such as edges and corners, through the designed DPEC module. To enable dynamic fusion of global and local features, a cross-attention feature fusion (CAFF) module is introduced, which adaptively aggregates global semantics and local details via a pixel-wise cross-attention mechanism. Experimental results on the NYUv2 and ScanNet datasets demonstrate that Swin-DPEC achieves superior performance compared with existing self-supervised methods, with relative errors of 0.123 and 0.138 and accuracy rates of 85.1% and 81.1%, respectively. Furthermore, the generated depth maps exhibit improved global structural accuracy and enhanced local detail fidelity.