UDepthDiff: Multi-Modal Physical Diffusion-Based Underwater Monocular Depth Estimation

Underwater monocular depth estimation is severely hindered by complex optical distortions, including water absorption, light scattering, and suspended particulates. To address these challenges, we propose UDepthDiff, a novel physics-guided framework that integrates multi-modal fusion with a diffusion model for robust depth prediction. Specifically, our method employs a conditional diffusion process in which multi-modal physical priors are extracted and injected into each denoising step to guide depth map generation. During the encoding stage, suspended particulates and spectral attenuation often obscure foreground-background depth relationships. To mitigate this, the Frequency-Enhanced Multi-modal Physics-Guided Generation (FEMPG) module utilizes Mamba-based scanning to reconstruct spectral content in the frequency domain, thereby generating saliency-informed physical priors. Subsequently, these multi-modal priors must be deeply integrated with depth features to ensure effective guidance for the diffusion model. The Attention-Weighted Multi-modal Fusion (AWMF) module facilitates this by aligning and fusing these priors with multi-scale depth features via self-attention mechanisms. In the decoding stage, a dual-branch diffusion decoder injects the fused guidance into each denoising step, preserving complementary local and global structures. Extensive experiments on the USOD10K and FLSea datasets demonstrate that UDepthDiff achieves state-of-the-art performance and significantly improves robustness against underwater optical distortions.