GOFish: fisheye-trained 3D Gaussian Splatting surpasses 180° FoV and extracts collision-ready meshes
janusch_patas · x · 2026-10-31
Amazon and University of Maryland researchers introduce GOFish, the first 3D Gaussian Splatting surface reconstruction method trained directly on fisheye images. It rasterizes fields of view beyond 180°, fixes 3DGUT's wide-angle failures (FoV cropping and depth-sort reversal past 90°), and extracts an aligned collision mesh from the same scene.
- Calibrated densification, mip filtering, and wide-angle appearance modeling handle pixel blur and lens falloff
- On the 200° FIORD and 205° FullCircle benchmarks, PSNR improves by 2.28 and 1.97 dB over the best prior method, with top mesh F-scores on FIORD and ScanNet++
- The result is a LiDAR-free, simulation-ready digital twin that renders in real time and traces the full FoV; code and models are open-sourced
More from Embodied
- China's dark warehouses run robots 24/7 — but do humanoids really fit? — luisdans · 2026-10-31
- Xiaomi's fabric-covered robotic hand shows off a coin-rolling finger trick — luisdans · 2026-10-31
- Will robots be solved by LLMs writing control code in real time? Debate sparks on X — lukas_m_ziegler · 2026-10-31
- Hy-Nomas: open-source interactive world model runs on consumer GPUs, rivals Genie 4 — bdsqlsz · 2026-10-31
- Tesla catches Waymo's Texas robotaxi fleet size in just 4 months vs 18 — mitchdeg · 2026-10-31
- AI toy spotted at a toy store in Ho Chi Minh City — "I'm the target customer, aren't I" — granawkins · 2026-10-31