An RGB-sonar fusion framework for underwater depth estimation enabled by cross-modal alignment

Underwater depth estimation is fundamental to numerous marine applications. Traditional depth sensors are costly and vulnerable to complex underwater optical conditions, while monocular vision-based methods, though flexible, suffer from inherent scale ambiguity that yields unreliable depth estimates. Multibeam sonar is unaffected by light attenuation and water turbidity, and can provide absolute ranging information to compensate for scale ambiguity. Therefore, we propose RSFNet, a supervised RGB-Sonar fusion framework for underwater depth estimation. To address the scarcity of datasets, we construct a synthetic dataset RSSUD and a real-world dataset RSRUD as training benchmarks. To address cross-modal misalignment caused by different imaging mechanisms, we design a Sonar Modality Alignment Module that transforms the sonar image into an STR image spatially aligned with RGB, providing coarse absolute distance priors to mitigate monocular ambiguity. The network employs a dual-branch encoder to extract semantic texture features from RGB and scale features from STR respectively and fuses them, achieving feature decoding through a multi-scale feature reconstruction decoder, followed by global context modeling via a Transformer-based module. Extensive experiments demonstrate that RSFNet significantly outperforms existing single-modal methods in depth estimation accuracy while maintaining model compactness. Ablation studies validate the effectiveness of our model.