Self-Supervised Monocular Depth Estimation With Progressive Enhancement of Local-to-Global Visual Perception

Self-supervised monocular depth estimation trains by utilizing the structure of the data itself without relying on ground-truth depth labels, gaining widespread attention in fields such as autonomous driving. However, many existing methods adopt the popular encoder-decoder structure, but this has deficiencies in refining both local and global visual clues, limiting its performance in detail recovery and spatial modeling. In this paper, we propose LGEDepth, a novel method to progressively enhance local to global visual perception. In LGEDepth, we design two crucial components to comprehensively refine local and global information after the encoder-decoder. The first is the patch-wise refinement tokenizer (PWRT), which fully refines the detailed information in local regions based on a local attention strategy with low performance overhead, effectively enhancing the local visual perception of the model. The second is the hierarchical interaction module (HIM), which better aggregates multi-scale contextual information through a cross-level interaction manner while determining the optimal depth bins that adapt to the depth distribution characteristics of the scene, effectively enhancing the global visual perception of the model. The experimental results on the KITTI, Cityscapes, Make3D and RUGD datasets demonstrate that LGEDepth achieves state-of-the-art performance and exhibits strong generalization ability, outperforming existing competitors.