You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This model is provided "AS IS" without warranty of any kind. The author is not liable for any damages or losses arising from the use of this model. By accessing this model, you agree to use it at your own risk.
Log in or Sign Up to review the conditions and access this model content.
NEMO β Underwater Metric Depth Estimation
Fine-tuned Depth Anything V2 for monocular metric depth estimation in underwater scenes.
Trained on ~1,010 real underwater RGB-depth pairs from Sea-thru (Red Sea) and SQUID (Red Sea + Mediterranean), using scale-invariant log loss with geometric augmentation (rotation Β±30Β°, random crop 70β100%).
Key findings
- Indoor init beats Outdoor β despite its 20 m depth ceiling, Indoor priors transfer better than Outdoor (80 m) because underwater visuals resemble close-range indoor geometry, not road scenes. Ξ΄β: 0.979 vs 0.947.
- Fine-tuning is essential β zero-shot Ξ΄β jumps from 0.197 to 0.979 after fine-tuning.
- Augmentation matters β rotation + crop significantly improves cross-dataset generalization (Satil Ξ΄β: 0.557 β 0.873).
Limitations
- Struggles on deep, high-backscatter scenes (D5 shipwreck, abs_rel > 1.8).
- Sea-thru test frames share dive sites with training data; SQUID cross-dataset results are more honest but based on very small sample counts.
- Downloads last month
- 8
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support