acceptodds
Under review as a conference paper at ICLR 2027

OC-GS: Gaussian Splatting for Irregular Turntable Capture

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 15.83dB, 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 7.88dB over keeping those estimates fixed. An ablation study shows that both image-derived angle initialization and the shared motion model contribute to the improvement. On real captures, OC-GS's refinement increases mean foreground PSNR by 0.70dB. Results show that refining uncertain angles within a shared motion model improves reconstruction from sparse, irregular turntable captures.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Gaussian Splatting, Turntable

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_xRx5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an interesting solution to the problem of uneven camera positions in 3DGS, but it addresses a very specific setting. The main idea is to estimate and refine the camera angles while reconstructing the object, constraining all camera movements to a physically consistent turntable rotation. This is based on the observation that viewing a rotating object is equivalent to moving a camera around a stationary object.

Critical strengths
  • Comprehensive experiments
Critical weaknesses
  • The method assumes a circular camera path with unequal angular gaps, rather than varying distances from the rotation axis. It is unclear how useful this specific setting is in practice. The method also requires foreground masks, known camera intrinsics, and ordered images.
  • VGGT already estimates camera poses without assuming equal angular spacing. Please clarify whether the main difficulty arises from irregular angular spacing itself, errors in sparse-view pose estimation, or converting VGGT-estimated poses into a shared turntable model. Reporting pose errors before circle fitting, after fitting, and after refinement would help distinguish these effects.
  • Refining camera angles can be detrimental when the initial angles are already accurate. The paper reports a 3.163.16 dB decrease in this experiment, which is consistent with overfitting.
Questions for authors

Please see weaknesses above.

Reviewer_PuP9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a Gaussian splatting framework for sparse turntable reconstruction with unknown, irregular angular spacing. It constrains cameras to a shared orbit with a common camera transform, rotation axis, and pivot, initializing geometry and angles from pretrained models via circle fitting, and then jointly refines the 3D scene, orbit parameters, and zero-mean per-view angular corrections. Evaluated on 100 Google Scanned Objects (6, 8, and 12 views) and real captures, it outperforms unconstrained pose-free baselines and uniform-spacing methods in foreground PSNR.

Critical strengths

Addressing irregular angular gaps caused by uneven manual/mechanical rotation or frame dropping in automated industrial/phenotyping pipelines is practically relevant. Constraining the optimization to a 1D circular orbit around a shared axis and pivot significantly reduces the degrees of freedom compared to unconstrained N x SE(3) camera bundle adjustment, reducing drift and deformation in sparse-view settings. The paper evaluates multiple initializers (VGGT, AnySplat, pi^3, Depth Anything 3, CUT3R, DUSt3R) and provides detailed ablations isolating the impact of initial angle quality, shared rig constraints, and joint angle refinement.

Critical weaknesses
  1. The framework relies almost entirely on high-capacity models like VGGT for usable initialization, collapsing in performance (by over 8 dB on Irregular-12) when initialized without network predictions or with weaker backbones.
  2. Because the 1D angular adjustment lacks regularization or uncertainty awareness, it fails under larger misalignments and actually degrades reconstruction quality by 3.1 dB when supplied with ground-truth starting poses. 3.The core contribution is an incremental 1D bundle adjustment constrained to a circular orbit, while the reported 16.2-minute runtime artificially excludes the essential, compute-heavy VGGT inference stage needed for fair comparison against feed-forward baselines.
Questions for authors

What is the total end-to-end wall-clock time and peak GPU memory footprint of OC-GS when fully accounting for VGGT feature extraction, camera/geometry inference, and circle fitting alongside per-scene optimization?

Does the optimization framework incorporate any confidence weighting or damping mechanism to prevent the 3.16 dB performance collapse observed when initial angles are already accurate, and what is the maximum angular initialization error from which the 1D refinement can reliably recover without converging to a degenerate local minimum?

Reviewer_5X43OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents OC-GS, an object-centric Gaussian Splatting framework for 3D reconstruction from sparse and irregularly spaced turntable captures. The primary objective is to reconstruct 3D objects and enable novel-view synthesis from a limited number of images with unknown rotation angles, assuming a fixed camera and an object rotating around a shared axis. OC-GS first leverages pretrained visual geometry models to estimate initial geometry and camera centers, from which nonuniform initial azimuth angles are derived. Subsequently, the method incorporates physical constraints of turntable motion to jointly optimize the Gaussian scene representation, shared camera pose, rotation axis, and other parameters. Experimental results demonstrate that OC-GS significantly outperforms existing pose-free Gaussian Splatting baselines under irregular capture conditions.

Critical strengths
  1. The paper addresses the practical problem of irregular turntable capture by introducing physically motivated motion constraints. This design effectively reduces the degrees of freedom in camera pose optimization while maintaining geometric consistency along the rotation trajectory.

  2. The paper provides comprehensive ablation studies under a unified training framework, comparing independent camera optimization, uniform angular spacing, frozen estimated angles, refined angles (OC-GS), and ground-truth angles. These controlled experiments provide stronger evidence supporting the effectiveness of the proposed approach.

Critical weaknesses
  1. The method relies on fitting a 3D circle to camera centers predicted by VGGT to estimate the initial rotation angles. This procedure may become unreliable or degenerate when input views are sparse or unevenly distributed, potentially limiting the robustness of the proposed approach.

  2. The method relies on accurate foreground masks and known camera intrinsics, raising concerns about the fairness of direct comparisons with external baselines. Moreover, the poor performance of certain baselines may not be solely attributable to the absence of per-view angular refinement, but could also result from unstable rotation-axis estimation, inappropriate scene bounds, or default configurations that are not well suited to the evaluated datasets.

  3. Compared with RotGS, the proposed approach does not appear to introduce a fundamentally new reconstruction or optimization mechanism. Instead, its contributions are primarily incremental, focusing on improved initialization and extensions to the parameterization of existing motion optimization frameworks.

  4. The main text provides a relatively brief description of the proposed method, with the methodological section occupying less than a page and containing only a limited number of core equations. Many important implementation and experimental details are deferred to the appendix, which reduces the clarity and self-contained nature of the main paper.

Questions for authors
  1. What are the fundamental differences between OC-GS and RotGS in terms of their optimization mechanisms? To what extent do the reported improvements stem from the strong geometric priors provided by VGGT, rather than the subsequent motion-constrained optimization? Specifically, what performance could be achieved by simply replacing the uniform-angle initialization in RotGS with VGGT-estimated angles and allowing independent per-view angular residual refinement?

  2. Could the authors provide comparisons under more consistent experimental configurations and adopt better-suited baseline settings, such as improved rotation-axis estimation and scene-boundary adjustments? Such comparisons would help establish whether the reported performance gains can be attributed to the proposed method rather than differences in baseline configurations.

  3. Could the authors provide experimental results on additional datasets? While angular refinement yields substantial improvements on synthetic data, the gains on real-world captures are relatively limited. Further experiments and analysis are needed to explain this discrepancy and assess the generalizability of the proposed approach.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.