OmniRay from JHU hits SOTA camera-agnostic monocular 3D reconstruction from one image
kwangmoo_yi · x · 2026-10-24
Researchers at Johns Hopkins and UVA present OmniRay, a feed-forward model that reconstructs 3D geometry from a single image across pinhole, fisheye, and panoramic cameras — no intrinsics or camera-type input needed at inference.
- Key idea: most monocular models assume a pinhole camera and distort on wide-angle imagery; the authors argue camera geometry has a hierarchy of complexity and the ray representation should mirror it.
- Composite ray fields: the ray module combines a monotone radial base (radially symmetric structure), a spherical-harmonic residual (smooth deviations), and a free-form residual (local variation), composed as successive rotations on the sphere.
- Output: per-pixel ray direction plus up-to-scale radial distance, from which a 3D pointcloud is recovered.
- Results: new state of the art in camera-agnostic monocular 3D reconstruction across camera configurations. Code and checkpoints are TBD.
More from Research
- TURTLE paper auto-evolves tool-use benchmarks, boosting hard tasks on AppWorld by 165% — MichaelD1729 · 2026-10-24
- Google's PLE benchmark: best AI agent scores only 50.6% on professional engineering deliverables — dair_ai · 2026-10-24
- Mathematician Danny Calegari: LLMs Mirror You Back, But Behave Better — fortnow · 2026-10-24
- No One Is Impressed by Good Results: Why AI-Proved Theorems Teach You Nothing — fortnow · 2026-10-24
- If Claude Had the Idea, Say So 'For Your Own Personal Soul': AI and Math Ethics — fortnow · 2026-10-24
- Top Mathematician Shows How He Uses Claude to Do Research and Build Talks — fortnow · 2026-10-24