Underwater Monocular Metric Depth Estimation: Real-World Benchmarks and Synthetic Fine-Tuning with Vision Foundation Models

1 min read

This paper benchmarks zero-shot and fine-tuned monocular metric depth models on real underwater datasets that have metric depth ground truth, including FLSea and SQUID, across water conditions and depth ranges. It then fine-tunes Depth Anything V2 (ViT-S encoder) on an underwater version of the synthetic Hypersim dataset, rendered with a physically based underwater image formation model.

Findings

  • Large models trained on terrestrial data, real or synthetic, do well in air but poorly underwater because of the domain shift (abstract).
  • The fine-tuned small model improves on every benchmark and beats the same model trained only on clean in-air Hypersim.

Limits

  • Abstract only so far. The fine-tuning water is simulated from a model, so it covers what the model covers (attenuation and backscatter) and not marine snow or moving light.

Relevance

It tests whether depth foundation models such as UniDepthV2 or MoGe-2 can be trusted underwater out of the box, and they cannot. A small ViT-S fine-tune is also the right size for a small robot. See underwater perception and kuan-2024-sea-thru-distance for how much depth error colour correction tolerates.