- Recent geometric foundation models have improved monocular depth estimation, but their benefits are limited for tiny models.
- DepthART is a compact model designed for robust on-device depth estimation across diverse scenes.
- It uses bias-resistant data sampling and camera-conditioned fine-tuning to address dataset-specific overfitting and unstable metric adaptation.
- These methods improve cross-dataset generalization and metric depth prediction in capacity-constrained models.
- DepthART has been accepted to ACM Multimedia 2026.