Paper Overview
Field: Computer Vision Authors: Jae Joong Lee, Bedrich Benes Published: 2026-09-25 arXiv: 2609.31572
Abstract
Uneven rotation and dropped frames make equal-angle assumptions unreliable for turntable reconstruction. We present OC-GS, an object-centric Gaussian splatting that refines each image's angle while maintaining a shared camera, rotation axis, and pivot. This orbit-consistent refinement jointly optimizes image-derived geometry and angles to reconstruct objects from sparse, irregular captures. On rendered objects with 12, 8, and 6 irregularly spaced views, OC-GS achieves mean foreground PSNR scores of 21.26, 19.36, and 15.83 dB, respectively, exceeding all four evaluated pose-free Gaussian splatting baselines in each condition. Under a shared trainer, refining image-estimated angles improves mean foreground PSNR by 7.88 dB over keeping those estimates fixed. An ablation study shows that both image-derived angle initialization and the shared motion model contribute to the gains. On real-world captures, OC-GS's refinement yields a 0.70 dB average foreground PSNR improvement. These results demonstrate that refining uncertain angles within a shared motion model improves reconstruction quality from sparse, irregular turntable captures.
Key Results
- Foreground PSNR on rendered objects: 21.26 dB (12 views), 19.36 dB (8 views), 15.83 dB (6 views)
- Outperforms all four evaluated pose-free Gaussian splatting baselines in each condition
- Refining image-estimated angles vs. fixing them: +7.88 dB mean foreground PSNR under a shared trainer
- Real captures: +0.70 dB average foreground PSNR from refinement