UniDepthV2 keeps the design of UniDepth (self-promptable camera module, pseudo-spherical output, geometric invariance loss) and changes three things: an edge-guided loss for sharper, better-placed depth edges, a simpler and faster architecture, and an extra output giving per-pixel uncertainty.
Findings
- The abstract reports better zero-shot performance and generalisation than prior methods across ten depth datasets. No headline number in the abstract.
Limits
- I read the abstract and README only.
- Whether the uncertainty output is calibrated out of domain is the thing I would want to check first, and nothing in the abstract says.
Relevance
A depth model that reports its own uncertainty is much easier to fuse safely: a map can down-weight pixels the model doubts instead of trusting every one. Underwater, the test is whether uncertainty rises with turbidity and range as it should, or stays flat while the depth goes wrong. See outlier-robust estimation and cai-2025-underwater-metric-depth.
Links
- arXiv · code
- Subfield: Perception