OC-GS: Gaussian Splatting for Irregular Turntable Capture

1 INTRODUCTION

• Uneven rotation and dropped frames make inter-image angles unequal, even when image order is known but the angles are not.Fewer images can reduce capture time, but complicate reconstruction.Source: §1 · p. 1
• Equal-angle schedules cannot represent arbitrary frame gaps, fixed inaccurate image-derived angles preserve errors, and independent camera updates can violate turntable motion.RotGS uses a frame-index schedule with an interpolated endpoint correction.Source: §1 · p. 1
• OC-GS jointly refines the Gaussian scene, shared camera, rotation axis, pivot, and per-view angle corrections, maintaining a shared trajectory while adjusting unequal gaps.It uses image predictions to initialize geometry and angles, without requiring measured angles or constant rotation speed.Source: §1 · p. 1, Figure 1 · p. 2
• Circle-fitting predicted camera centers initializes non-uniform angles, while subtracting the mean per-view correction adjusts relative spacing without adding a redundant common rotation.The image-derived geometry initializes the scene; shape and motion are then optimized together under the turntable constraint.Source: §1 · p. 2

2 OC-GS: ORBIT-CONSISTENT REFINEMENT

• OC-GS refines image-derived geometry and per-view angles together while sharing one camera, rotation axis, and pivot across views.The task assumes input RGB images, foreground masks, and camera intrinsics; segmentation and intrinsic calibration are not evaluated.Source: §2 · p. 2, §2.1 · p. 2

2.1 TURNTABLE RIG

• The turntable rig models each view as rotation about a pivot followed by a shared camera transform, grouping effective cameras into one trajectory with unequal gaps.The pivot is translated to the origin for rotation and then translated back; Gaussian centers and local frames share the rigid motion.Source: §2.1 · p. 2

2.2 ANGLE SOURCES AND SPACING ERRORS

• Synthetic initialization uses VGGT geometry and camera centers to fit angles; real captures use AnySplat-derived angles, and the initializer is held fixed across synthetic variants.Per-image corrections are mean-centered to adjust relative spacing, then optimized with the scene, camera, axis, and pivot while intrinsics remain fixed.Source: §2.2 · p. 3
• The component comparison contrasts independent cameras, shared-rig uniform angles, fixed image-derived angles, refined OC-GS angles, and a true-angle reference.Capture conditions include regular 12-view and irregular 12-, 8-, and 6-view settings; quality is evaluated on 36 target views using FG-PSNR, LPIPS, IoU, and relative rotation error.Source: §2.3 · p. 3, Figure 2 · p. 3

3.1 COMPARISON WITH EXISTING RECONSTRUCTION METHODS

• With 12, 8, and 6 irregular views, OC-GS achieved mean foreground PSNR scores of 21.26, 19.36, and 15.83dB, outperforming all evaluated pose-free GS methods at each view count.Quality decreased with fewer views, but OC-GS remained ahead of the baselines.Source: §3.1 · p. 4
• Selected comparisons show OC-GS preserving the Hereford Bull’s leg positions and silhouette and reconstructing the Bathroom Set’s main parts, while some baselines distort, displace, or omit content.The figure shows the same target view and crop across methods; across valid aligned cases, OC-GS averaged 1.43dB higher FG-PSNR and 0.103 lower object-crop LPIPS than Splatfacto with true cameras.Source: §3.1 · p. 4, Figure 3 · p. 4

3.2 WHICH COMPONENTS ACCOUNT FOR THE IMPROVEMENT?

• With fixed estimated angles, the shared rig reached 11.99dB mean FG-PSNR versus 12.42dB for independent cameras, while uniform angles under regular capture reached 26.29dB versus 12.04dB for frozen image-derived angles.Uniform-angle initialization plus refinement improved FG-PSNR by 1.96dB but remained 5.80dB below OC-GS; tenfold independent-camera learning rates added only 0.23dB, leaving a 6.92dB gap. The refinement gain remained 6.9–8.1dB under alternate evaluation choices, and refinement closed 9.08dB of the 18.68dB gap to the true-angle reference with twelve irregular views.Source: §3.2 · pp. 4–6
• Joint angle refinement raised mean FG-PSNR to 19.88dB, a 7.88dB improvement over frozen estimates, while lowering LPIPS from 0.281 to 0.142 and raising IoU from 0.718 to 0.860.Angular-spacing RMSE fell from 10.10° to 8.53°; Table 1 reports FG-PSNR variants under the same trainer and VGGT predictions.Source: §3.2 · p. 5, Table 1 · p. 5, Figure 4 · p. 5

3.3 REAL-CAPTURE EVIDENCE AND INITIALIZATION

• On real RotGS captures, optimizing AnySplat-derived angles improved mean FG-PSNR by 0.70dB and lowered mean LPIPS at every tested view count.VGGT was the best initializer on synthetic objects, 0.86dB above VGGT-Ω; AnySplat was best on the RotGS dataset, 0.36dB above π3.Source: §3.3 · p. 6, Table 3 · p. 6

3.4 WHEN DOES ANGULAR REFINEMENT HELP?

• With noisy angles, refinement improved FG-PSNR at every tested RMS error from 0.5° to 10°, with gains rising from 4.75dB to 11.25dB; with exact angles, it worsened unseen-view results.At 10° RMS, FG-PSNR rose from 9.08dB to 20.33dB; with exact angles, the variants fit training images similarly well.Source: §3.4 · p. 7, Table 5 · p. 7

3.5 SOURCES OF ERROR IN THE BASELINE RECONSTRUCTIONS

• For RotGS, fixing the axis to its true value raised mean FG-PSNR from 15.97 to 29.45dB; image-based baselines showed camera-scene inconsistency or limited training fit, while Nerfacto’s bounded field raised target FG-PSNR from 6.80 to 17.78dB.These diagnostic results cover separate baseline experiments and settings.Source: §3.5 · pp. 7–8
• OC-GS used the least peak GPU memory at 1.51 GB and fit in 16.2 minutes, less time than RotGS and Nerfacto; InstantSplat and feed-forward methods were faster but had lower reconstruction quality.VGGT inference was precomputed for OC-GS; Table 6 reports quality on valid benchmark runs and cost from separate profiling runs.Source: §3.5 · p. 8, Table 6 · p. 8

4 LIMITATIONS

• OC-GS assumes a fixed axis with minimal axis noise and depends on VGGT initialization.The authors identify adapting refinement to initialization quality as future work.Source: §4 · p. 8

5 CONCLUSION

• OC-GS jointly optimizes image-derived angles and Gaussian geometry under a shared motion model to reconstruct sparse, irregular turntable images.Source: §5 · p. 8
• OC-GS outperformed the evaluated pose-free GS baselines in all three irregular capture conditions, while angle refinement added 9.08 dB FG-PSNR at twelve views and 4.22 dB at six versus fixed estimated angles.Source: §5 · p. 8
• Under the same trainer and initialization, refining image-estimated angles improved mean foreground PSNR by 7.88dB over keeping those estimates fixed.Source: §5 · p. 8
• The authors conclude that refining uncertain angles within a shared motion model improves reconstruction from sparse, irregular turntable captures.Source: §5 · p. 8

• Pose-free Gaussian methods reconstruct scenes from images, while OC-GS constrains views to one camera, axis, and pivot with a separate angle for each image.Source: §6 · p. 8
• Geometry-and-camera predictors provide estimates that OC-GS converts into initial turntable angles and refines using the captured images.Source: §6 · pp. 8–9
• Unlike general joint camera-scene refinement methods, OC-GS constrains camera updates to a shared rotation trajectory throughout optimization.Source: §6 · p. 9
• Earlier turntable methods used silhouettes, circular-motion constraints, or conic fitting; RotGS uses a uniform index schedule that cannot represent arbitrary frame gaps.Source: §6 · p. 9
• OC-GS initializes nonuniform angles from images and refines each angle around a shared axis.Source: §6 · p. 9

A BENCHMARK AND IMPLEMENTATION DETAILS

• The synthetic benchmark uses 100 Google Scanned Objects rendered at 512 × 512 pixels, with 36 calibrated target views per capture condition.Real captures come from the dataset released with RotGS.Source: A BENCHMARK AND IMPLEMENTATION DETAILS · p. 12
• Irregular-view gaps are sampled from a lognormal distribution, smoothed, and rescaled to 360°; sparse subsets use rounded-down indices from a 24-image sequence.The sparse subsets are not nested, and targets within 0.05° of a training view are excluded.Source: A BENCHMARK AND IMPLEMENTATION DETAILS · p. 12
• Primary component experiments use the same VGGT predictions, 30k training steps, and pure L1 photometric loss, with silhouette, depth, consistency, normal, and distortion losses disabled.Angle refinement uses a 0.001 learning rate, mean-correction subtraction, and no angle penalty.Source: A BENCHMARK AND IMPLEMENTATION DETAILS · p. 12
• Initial angles come from a plane-and-circle fit to all predicted camera centers, with the first image defining zero rotation and no outlier rejection.Refinement does not enforce the initial angle ordering; primary component experiments and optimized external comparisons use three seeds.Source: A BENCHMARK AND IMPLEMENTATION DETAILS · p. 12

B CAMERA ALIGNMENT AND EVALUATION

• Camera alignment fits rotation from training-camera orientations, then fits scale and translation to camera centers by least squares and applies one transform to all target cameras.Umeyama’s camera-center fit is used as a cross-check; flagged disagreements do not alter the main scores.Source: B CAMERA ALIGNMENT AND EVALUATION · p. 12
• 8.05dB is the refinement gain when evaluation is restricted to orientation-based alignments stable across all variants and seeds.The alignment flags do not change the main scores.Source: B CAMERA ALIGNMENT AND EVALUATION · p. 12
• FG-PSNR measures RGB error on target foreground pixels, while LPIPS, silhouette IoU, and RRE measure perceptual similarity, mask overlap, and relative-rotation discrepancy.Object-crop LPIPS uses a padded target-mask bounding box, and IoU thresholds rendered opacity and target masks at 0.5.Source: B CAMERA ALIGNMENT AND EVALUATION · p. 12

C EXTERNAL BASELINES AND COMPLETE COMPARISONS

• External comparisons include feed-forward methods and calibrated references, while the complete comparison averages methods over their valid cases and paired differences use shared valid cases.Calibrated references receive true cameras; paired FG-PSNR and object-LPIPS differences favor OC-GS when positive and negative, respectively.Source: C EXTERNAL BASELINES AND COMPLETE… · p. 13, Table A2 · p. 13, Table A3 · p. 13
• 30k-step mean FG-PSNR is 15.74dB in the benchmark and 15.42dB in separate profiling runs, with a 0.51dB mean absolute paired difference.Runtime and peak memory are profiled separately; the cost benchmark uses an RTX A5000, and calibrated-reference costs are reported separately.Source: C EXTERNAL BASELINES AND COMPLETE… · pp. 13–14, Table A5 · p. 14, Table A6 · p. 14

D REAL-CAPTURE EXPERIMENTS

• The experiment evaluates nine captures at 24, 17, 13, and 7 training views using 30k steps and a pure L1 loss.Target cameras are estimated rather than independently calibrated.Source: D REAL-CAPTURE EXPERIMENTS · p. 14
• Angle refinement is the only difference between the fixed-angle and refined-angle variants; both use the same AnySplat angles, input images, target views, and camera conventions.The table reports means across nine captures, with FG-PSNR and paired gain in dB and LPIPS listed frozen then refined.Source: D REAL-CAPTURE EXPERIMENTS · p. 14, Table A7 · p. 14
• Target cameras are estimated rather than independently calibrated, so the comparison evaluates reconstruction under estimated target-camera poses.This setting applies to the nine-capture experiment.Source: D REAL-CAPTURE EXPERIMENTS · p. 14

E SENSITIVITY AND COMPONENT EXPERIMENTS

• Under regular-12 capture with fixed VGGT predictions, the transition from harmful to helpful refinement falls between 0.1° and 0.2° angle error.The supplied angles also affect initialization, so comparisons across error levels include that effect.Source: E SENSITIVITY AND COMPONENT EXPERIMENTS · p. 14, Table A8 · p. 14
• With exact initial angles, refinement reduces target-view quality despite similar training photometric errors.Source: E SENSITIVITY AND COMPONENT EXPERIMENTS · p. 15
• Three-seed mean FG-PSNR gains are 7.68dB with pure L1, 5.99dB with 0.8/0.2 L1/SSIM, and 4.88dB with 0.6/0.4, with every capture condition improving.Tests with 0.9/0.1, 0.4/0.6, and pure SSIM also show positive gains.Source: E SENSITIVITY AND COMPONENT EXPERIMENTS · p. 15
• Refinement gains remain 7.16dB and 6.91dB when target views within 5° and 10° of training angles are excluded, respectively.The benefit persists when evaluation excludes views close to the inputs.Source: E SENSITIVITY AND COMPONENT EXPERIMENTS · p. 15

F.1 ROTGS: CAMERA FIT AND AXIS SENSITIVITY

• RotGS reaches 33.35dB FG-PSNR on regular-12 training views using estimated cameras but 14.24dB using aligned reference cameras.The repeated endpoint image is excluded.Source: §F.1 · p. 15
• Fixing the axis to its true value raises mean FG-PSNR from 15.97 to 29.45dB, with every tested object improving.The initial geometry and 30k-step budget are fixed while the pivot remains optimizable.Source: §F.1 · p. 15

F.2 IMAGE-BASED BASELINES: CAMERA CONSISTENCY

• Camera replacement lowers training-image quality for AnySplat and InstantSplat, while SPFSplat already fits training images poorly before replacement.Source: Table A9 · p. 15, §F.2 · p. 15
• Correcting InstantSplat’s exported vertical focal length raises target FG-PSNR from 5.53 to 5.87dB, but explains only a small part of OC-GS’s gap.Source: §F.2 · p. 15
• Nerfacto can fit sparse training images without recovering correct object geometry; density outside the object region limits visual quality from unseen viewpoints.Source: §F.2 · p. 16