USEDepth: Semantic-Conditioned Monocular Depth Sensing for Underwater Perception

Abstract

Robust monocular depth sensing is critical for autonomous underwater vehicles (AUVs), yet underwater optical degradation and interscene metric ambiguity weaken the appearance cues used by conventional image-to-depth models. Existing methods mainly improve local visual robustness but rarely exploit environmental context to guide depth representation learning across varying underwater conditions. We present USEDepth, a descriptor-guided framework that uses structured environmental context to condition representation learning for underwater depth estimation. The context descriptor encodes substrate type, biota density, and lighting condition, three coarse factors that are consistently recoverable from underwater imagery and associated with scene-level depth ambiguity. These descriptors are distilled from 1137 underwater captions, encoded offline with a frozen CLIP text encoder, and integrated through descriptor attention and channelwise semantic modulation (CSEM) to guide the learned visual representation without runtime vision–language inference, dense labels, auxiliary sensors, or multiview geometry. On FLSea, USEDepth achieves 0.110 absolute relative (Abs Rel) and improves near-field accuracy by 30.8% for depths below 3 m. Without target-domain fine-tuning, it achieves 0.210 Abs Rel on SQUID, indicating improved transfer across underwater conditions. The model runs at 46 FPS on an RTX 3070 and 31 FPS on a Jetson AGX Orin. Ablations show that shuffled descriptors increase error by 168.5%, whereas removing descriptors after convergence causes only a 1.7% degradation, supporting the interpretation that environmental context primarily acts as a training-time conditioning signal that shapes the learned visual representation rather than serving as a direct depth cue during inference.