Turntable 3D scanning just got more forgiving of shaky cameras and skipped frames.
Researchers built OC-GS, an object-centric Gaussian splatting method for objects filmed on a rotating turntable that doesn't require perfectly even rotation speed or a complete set of frames. Rather than trusting the equal-angle assumption most turntable pipelines rely on, it estimates each image's rotation angle from the image itself, then refines those estimates while keeping one shared camera, rotation axis, and pivot point across the whole capture. On synthetic test objects filmed from 12, 8, and 6 irregularly spaced viewpoints, OC-GS beat four existing pose-free Gaussian splatting baselines, scoring average foreground image-quality (PSNR) of 21.26, 19.36, and 15.83 decibels respectively. Refining the angle estimates rather than locking them in place added a 7.88 dB average quality boost under a shared trainer, and the approach still helped on real, non-synthetic captures, adding 0.70 dB.
Most 3D-scanning pipelines assume a tidy setup: constant motor speed, every frame captured, camera never wobbling. Cheap consumer turntable rigs rarely deliver that. Treating rotation angles as estimates to refine, instead of fixed inputs, could make usable 3D object scans possible with far less hardware precision - relevant for anyone building 3D catalogs, AR assets, or game props without a studio-grade rig.
The paper only tests turntable setups, so this won't replace general structure-from-motion pipelines - it's a fix for one common, specific capture headache, not a broader reconstruction breakthrough.