Wang et al. (2025) — ARXIV-2503.11651
Key takeaways
- One network, all the attributes. Camera parameters, point maps, depth maps and 3D point tracks, inferred directly and together, “from one, a few, or hundreds of its views”.
- The framing claim: 3D vision models “have typically been constrained to and specialized for single tasks”, and that specialisation is what this removes. The tasks were separate because the pipeline was, not because the problems are.
- Under one second, and still “outperforming alternatives that require post-processing with visual geometry optimization techniques”. Beating optimisation with a forward pass is the strong form: the iterative refinement that justified the classical pipeline is not buying what it was thought to.
- State of the art across camera parameter estimation, multi-view depth, dense point cloud reconstruction and 3D point tracking — one model, four tasks that used to be four literatures.
- Useful as a backbone, improving non-rigid point tracking and feed-forward novel view synthesis, which is the tell that it learned geometry rather than four separate output heads.
Standing in the anthology
Sources SOTA-236 with LIT-385. The two are a line rather than a duplicate: DUSt3R establishes that reconstruction without calibration or pose is possible and still needs a global alignment step for more than two images; this one removes the remaining optimisation and scales the view count, which is what turns a paradigm claim into a default.
Two independent groups is also what raises the practice above one result — the author lists share a surname and nothing else.
For this record the interesting consequence is the backbone finding. It puts geometry estimation in the same category as the pretrained encoders the rest of the corpus takes for granted, which is the point at which a subject stops being a separate field and becomes a component.