CNN-Transformer Feature Aggregation for Underwater Self-Supervised Multiframe Monocular Depth Estimation

Depth estimation plays a critical role in perceiving and understanding underwater scenes. However, accurate underwater depth estimation remains challenging due to nonuniform degradations, including color distortion, blur, and low contrast. To address these challenges, a dual-component framework is proposed, which integrates a CNN-based depth prediction component and a transformer-based image reconstruction component. The depth prediction component extracts depth feature from consecutive frames and estimates depth. Meanwhile, the image reconstruction component employs random masking to simulate degradations, compelling the component to predict masked regions using contextual cues, providing supplementary information to enhance depth prediction. A hybrid feature fusion module is designed to integrate local depth features and global contextual information from both components, ensuring greater robustness and accuracy. Additionally, global contextual information is embedded into the depth prediction component at multiple encoding stages, enabling fine-grained depth estimation. Experimental results on the FLSea and USOD10K datasets demonstrate that CTFDepth outperforms previous methods across all evaluation metrics, highlighting its effectiveness in underwater depth estimation. The source code in this work is available at https://github.com/caremore/CTFDepth