Any3R: one feedforward model handles 13 camera types for 3D reconstruction, beating OmniVGGT and PanoVGGT
zhenjun_zhao · x · 2026-10-29
The paper "Any3R: Feedforward 3D Reconstruction from Any Camera Type" (arXiv 2610.36150) introduces a feedforward model that accepts any central camera — pinhole, wide fisheye, panorama — and any mix of them per scene, with no per-lens retraining.
Method:
- Existing geometry foundation models like VGGT assume pinhole images; positional encoding and training loss both bind an image position to a direction in space, depending on the lens;
- Any3R conditions both on the dense ray field (unit direction through every pixel), which describes any central camera without lens type, distortion coefficients or FOV bounds;
- Angular coordinates feed each patch's measured ray angle into rotary encoding per view, and solid-angle weighting scales the point loss by each pixel's spherical area, decoupling encoding and supervision from the lens.
Results: on a benchmark of 13 central cameras across 3 datasets (each method fusing views with its own predicted poses), Any3R is the most consistently accurate method, beating OmniVGGT, PanoVGGT, Wid3R, Argus and MapA-ray on Acc↓ and AUC@15°/30° (e.g. indoor ERP AUC@30° 96.8 vs PanoVGGT's 93.2). It enables scalable 3D reconstruction from heterogeneous crowdsourced and archival imagery.
More from Research
- Information-theoretic scaling laws predict memorization and FID in diffusion models — LucaAmb · 2026-10-29
- ICLR reviewer hands out scores of 1,1,1,1 and 2 across five papers — chhaviyadav_ · 2026-10-29
- kalomaze calls out hnet baseline: effective depth unmatched, no wall clock reported — kalomaze · 2026-10-29
- AVERT Study Launched: Can Autonomous Vehicles Reduce Trauma? — zakkohane · 2026-10-29
- Bad binder design drove only ~3% of Phase 2/3 failures — the rest elude AlphaFold — clarejtbirch · 2026-10-29
- Demo shows emergent 'in-context tokenization' beating fixed tokenizers — willdepue · 2026-10-29