Contextual information-based amalgamated CNN-Transformer network for self-supervised monocular depth estimation

Estimating depth from images is a crucial computer vision task with wide-ranging applications in fields such as autonomous driving, drones, and virtual reality. Self-supervised monocular depth estimation utilizes image sequences to achieve semi-supervised learning and has shown promising application prospects. However, current self-supervised methods still suffer from deficiencies in enhancing feature dependencies and properly handling local information, resulting in limited performance and low prediction accuracy. In this work, we propose a novel network architecture, ACTADepth, based on the U-Net framework, aimed at further improving prediction accuracy. The network utilizes anamalgamated CNN-Transformer as the depth encoder to capture and convey contextual information. Meanwhile, an efficient decoder is designed to effectively exploit both global and local cues in images. In the proposed channel attention module, dependencies between any two channel mappings are captured to enhance feature representation. By establishing connected paths between multi-scale local features and the global decoding stream via the feature fusion module, the network can integrate both representations and recover depth details effectively. The proposed network demonstrates competitive results on the KITTI dataset.