Monocular Depth Estimation Algorithm Guided by Global Structure and Local Details

In the current encoder-decoder architecture for monocular depth estimation, there are issues with unclear structural details of depth scene objects during the decoder's upsampling process, as well as insufficient capability to extract contextual information from the scene. To address these problems, this paper proposes a monocular depth estimation algorithm guided by global structure and local details. The algorithm utilizes a pre-trained Swin-Transformer as the encoder to downsample the monocular image and extract multi-scale features from different levels of encoder representations. During the decoder phase, high-level features are fused with low-level features to create feature guidance maps enriched with structural semantic and positional information. These maps are then refined layer by layer with local detail feature maps output from a context aggregation module. This process ensures the model retains more global and local information during the multi-scale feature fusion. Finally, global structure guidance maps and local detail guidance maps from all layers are combined to predict the final depth map. Extensive experimental results on both indoor and outdoor datasets demonstrate the superiority of the proposed model. Compared to the AdaBins algorithm, our approach reduces RMSE by 10.46% and 7.41% on the KITTI and NYU Depth v2 datasets, respectively.