Depth-PVT: Pyramid Vision Transformer with Channel Attention for Depth Estimation

One of the most important machine vision tasks is monocular depth prediction, essential for applications such as autonomous navigation and augmented reality. While CNN-based architectures are widely used, they are limited by restricted receptive fields, which hinder effective multi-scale feature extraction. In order to overcome these issues, we develop Depth-PVT, an innovative architecture that integrates Efficient Channel Attention (ECA) with the Pyramid Vision Transformer (PVT). This integration enhances feature representation by emphasizing salient channel-wise features and suppressing irrelevant ones with minimal computational overhead. Depth-PVT incorporates three key modules within the PVT framework: the cascaded fusion module (CFM) for aggregating semantic and spatial information of high level features, the camouflage identification module (CIM) for capturing complex depth cues and low level features, and the similarity aggregation module (SAM) for fusing cross-scale features, thereby enriching depth details with high-level semantics. , Depth-PVT is trained on the NYU-Depth V2 dataset and finally surpasses current cutting-edge techniques, attaining better results in metrics like absolute relative error, RMSE, and RMSE(log).