UniDepth predicts metric 3D points from a single image with no camera intrinsics given. A self-promptable camera module predicts a dense camera representation that conditions the depth features. The output is pseudo-spherical, which separates camera from depth, and a geometric invariance loss pushes the camera-conditioned depth features to stay the same across augmentations of one scene.
Findings
- The abstract reports better zero-shot results than prior methods across ten datasets, even against methods trained on the test domains. No headline number in the abstract.
Limits
- I read the abstract and README only.
- Metric scale and intrinsics both come from learned priors, so both can be wrong together without any visible sign.
- Ten in-air datasets. cai-2025-underwater-metric-depth finds models of this kind lose much of their accuracy underwater.
Relevance
Recovering intrinsics from the image is useful when a housing or a swapped lens changes the camera. For a team of robots, the open question is how to fuse metric depth from different robots whose learned scales disagree. See UniDepthV2 and foundation models for robotics.
Links
- arXiv · code
- Subfield: Perception