acceptodds
Under review as a conference paper at ICLR 2027

Proxyboros: Motion-Faithful Dynamic Self-Intersection Repair for Rigged Mesh Animations

Abstract

Physically plausible animation of rigged shapes is essential for realistic and high-quality 3D content creation. However, motion sequences produced by existing generation models often exhibit surface collisions and self-intersections, resulting in physically implausible geometry and noticeable visual artifacts. Resolving these invalid surface geometries while faithfully preserving the spatiotemporal characteristics of the original motion remains a challenging yet under-explored problem. To address this, we propose Proxyboros, an optimization framework for dynamic self-intersection repair that alternates mesh-space repair with rig-space pose adaptation to progressively guide the evolving rig toward temporally coherent skeletal motion. Within each step, Mesh-Space Proxy Exploration first extracts intermediate geometric proxies that capture local repair tendencies before independent repairs crystallize into temporally inconsistent static solutions. Proxy-Guided Pose Adaptation then incorporates this guidance into a temporally coupled skeletal trajectory using surface-aware proxy guidance, absorbing stable surface-separation cues while relaxing ambiguous local displacements. Experiments on SMPL-H sequences and diverse non-SMPL rigged models demonstrate favorable overall performance over state-of-the-art methods in repair quality, pose preservation, and motion fidelity, with demonstrated applicability across heterogeneous mesh and skeleton topologies. See our anonymous project page for more video demos: https://anonymousiclr2724.github.io/Proxyboros.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Shape modeling, self-intersection repair, self-collision repair, motion generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_tsGEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Proxyboros repairs self-intersections in rigged animations by alternating mesh-space collision repair with rig-space fitting. Temporally parameterized pose corrections aim to preserve the original motion while maintaining compatibility with the input rig. Evaluation includes collision metrics, motion-preservation measures, ablations, and a user study.

Critical strengths

The paper demonstrates a useful collision–motion trade-off. Compared with ISIR, collision reduction is broadly comparable, while motion-direction distortion is substantially lower. This directly addresses a practical limitation of collision repair that alters the intended animation.

The alternating formulation has meaningful experimental support. Comparisons with one-shot repair followed by rig fitting, including matched iteration budgets, and temporal ablations provide evidence that the proposed workflow improves over straightforward sequential processing.

Critical weaknesses

The contribution is not fully separated from the temporal representation and regularization. A direct rig-space collision optimizer using the same spline parameterization, pose regularization, and comparable budget would be a stronger control. Without it, the necessity of alternating proxies remains partly unresolved.

Collision-metric robustness and cross-rig generalization need stronger evidence. PVR depends on semantic partitioning and voxelization, which can omit within-part intersections and introduce resolution sensitivity. This matters because its advantage over ISIR is only 0.40 percentage points, while IFR is slightly lower. Quantitative evidence on non-SMPL rigs is also limited.

Questions for authors

Does the method still outperform direct rig-space collision optimization when temporal parameterization, regularization, and computational budget are matched?

Are the conclusions stable across voxel resolutions and an independent collision measure that includes within-part intersections?

What are the quantitative collision and motion-preservation results for the non-SMPL characters?

1 comment

We thank the reviewer for the positive feedback on the practical value of our work. Below, we respond to the main concerns.

Q1:

Performance over direct rig-space optimization. To further validate the effectiveness of PMRO, we introduce a variant that directly optimizes the mesh-space repair objective in rig space under a matched optimization budget, removing the progressive mesh–rig interaction while retaining the same temporal parameterization and pose regularization. The quantitative results are shown below.

MethodPVR (%) ↑IFR (%) ↑SRS (%) ↑VD (×10²) ↓VAD (×10³) ↓VAM (×10³) ↓
w/o PMRO95.0892.9990.3311.2321.3692.631
Full94.6489.7393.465.6300.6072.202

Although direct rig-space optimization achieves higher collision reduction, it substantially degrades pose preservation and motion fidelity. This confirms the kinematic coupling issue illustrated in Figure 1 of the main paper, where local collision corrections induce excessive skeletal articulation and unintended motion changes. By progressively incorporating mesh-space repair guidance into rig-space optimization, PMRO achieves a substantially better balance between collision resolution and motion preservation.

Q2:

Voxel resolution sensitivity. We would like to clarify that voxel resolution affects the numerical estimation of PVR, rather than the repair results themselves. We re-evaluate all methods at different voxel resolutions, and the results are shown below.

PVR Voxel Res.64128160256512
ISIR95.3294.0294.2494.3394.39
Mesh-Utg94.8793.4893.5493.2493.42
ISIR + skeleton fitting93.3593.1793.2793.3093.35
Mesh-Utg + skeleton fitting89.8989.5789.6389.6589.68
PoseShield91.9392.0492.0392.1492.16
Ours98.3498.0294.6494.7794.80

At coarse resolutions, small residual penetrations may be missed by coarse voxelization, leading to higher PVR values.
However, our method consistently achieves the highest PVR across all resolutions. More importantly, its advantage over the other baselines remains stable at resolutions above 160, showing that our reported improvement does not depend on a favorable voxel resolution.

Within-part intersections. We clarify that within-part intersections are already included in our evaluation through IFR, which measures triangle-level intersections on the complete mesh independently of semantic partitioning. PVR instead measures volumetric penetration between body parts, where finer partitioning does not necessarily yield more accurate estimates. To examine this, we additionally evaluate PVR using a finer bone-level partition of SMPL-H with the same skinning-weight-based segmentation and surface capping procedure.

Part-levelBone-level
ISIR + skeleton fitting93.2782.65
Mesh-Utg + skeleton fitting89.6380.06
PoseShield92.0381.28
Ours94.6483.26

While all methods exhibit substantially lower PVR, ours remains the best. This decrease mainly arises from artificial overlaps near bone boundaries, where non-rigid skinning deformations cause adjacent capped segments to overlap without actual surface intersections. Our original 13-part partition avoids such artifacts while capturing major inter-part penetrations. Moreover, finger articulation remains fixed in the evaluated PoseShield sequences, supporting the grouping of each hand into a single part.

Q3:

Non-SMPL quantitative evaluation. We have added quantitative results on 30 non-SMPL animations from Objaverse and Truebone Zoo, covering diverse geometries and skeletal structures. As shown in Table 2 of the revised manuscript, our method achieves competitive repair quality with strong pose preservation and the best motion fidelity, further supporting its generalizability beyond SMPL-H.

MethodsIFR (%)↑SRS (%)↑VD↓ (×10²)VAD↓ (×10³)VAM↓ (×10³)
ISIR + skeleton fit81.4989.491.4634.97712.278
Mesh-Utg + skeleton fit87.3878.304.69317.52423.152
Ours84.5689.131.4311.72510.343
Reviewer_qxFYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Proxyboros, an optimization framework for repairing self-intersections in rigged mesh animations while preserving the original motion. Its main contribution is an alternating mesh–rig optimization scheme: short mesh-space repair steps generate intermediate proxies, which guide temporally coupled skeletal pose updates through cubic B-spline corrections, surface-aware alignment, and pose regularization. Validation includes comparisons and ablations on 100 SMPL-H sequences, alongside qualitative results and a user study involving non-SMPL humanoids and animals. The reported results support a favorable balance between penetration reduction, pose preservation, and temporal motion fidelity.

Critical strengths

A well-motivated framework for reconciling local collision repair with skeletal motion preservation. Intermediate mesh proxies allow local repair directions to be progressively incorporated into temporally coherent skeletal motion. The equal-update-budget ablation supports this alternating design: compared with one-shot repair followed by adaptation, it improves skeleton rotation similarity from 89.73% to 93.46% and reduces acceleration discrepancy from 0.863 to 0.607, with a modest decrease in repair quality. This provides evidence that progressive interaction improves the repair–preservation trade-off. Meaningful improvements in motion fidelity while maintaining strong repair performance. On the SMPL-H benchmark, Proxyboros achieves 94.64% penetration-volume reduction and an acceleration discrepancy of 0.607, compared with 92.03% and 3.071 for PoseShield. These results support the practical value of the method: it substantially reduces penetration while better retaining the input motion’s temporal characteristics, a central requirement for animation repair.

Critical weaknesses

Insufficient objective evaluation on non-SMPL models. Applicability across heterogeneous mesh and skeleton topologies is a central claim, yet the non-SMPL results primarily consist of qualitative examples and subjective user preferences. Objective metrics such as IFR, SRS, VD, and VAD are not reported for these models, although they could be evaluated with appropriate conventions. This leaves the extent to which the repair quality and motion fidelity observed on SMPL-H transfer to other rigs insufficiently established. Limited characterization of experimental coverage and scalability. Although the main benchmark contains 100 SMPL-H sequences, their length distribution is not reported, and runtime is evaluated on a single 90-frame sequence without a sequence-length scaling study. The non-SMPL user study includes 10 humanoid cases and five animal cases, while the size and composition of the full non-SMPL test set remain unclear. These gaps limit assessment of performance on longer animations and across a broader range of rig and mesh complexities.

Questions for authors

Non-SMPL quantitative evaluation. Could you report objective repair-quality and motion-fidelity metrics for non-SMPL models, with results broken down by humanoid type, animal species, and challenging mesh configurations such as disconnected components or open boundaries? This would clarify whether the benefits are consistent across different skeletal structures and mesh conditions. Evaluation coverage and sequence-length scalability. Could you clarify the size and diversity of the full test set, including sequence-length statistics and the numbers of distinct non-SMPL assets and motions? Beyond the single 90-frame runtime example, how do runtime, peak memory, repair quality, and motion fidelity change for longer sequences, and are these optimized jointly or in temporal windows?

3 comments

We thank the reviewer for the positive feedback on the motivation and practical value of our method. Below, we respond to the main concerns.

Q1:

Non-SMPL quantitative evaluation. We have added quantitative results on 30 non-SMPL animations from Objaverse and Truebone Zoo, covering diverse geometries and skeletal structures. As shown in Table 2 of the revised manuscript, our method achieves competitive repair quality with strong pose preservation and the best motion fidelity, further supporting its generalizability beyond SMPL-H.

MethodsIFR (%)↑SRS (%)↑VD↓ (×10²)VAD↓ (×10³)VAM↓ (×10³)
ISIR + skeleton fit81.4989.491.4634.97712.278
Mesh-Utg + skeleton fit87.3878.304.69317.52423.152
Ours84.5689.131.4311.72510.343

Q2:

Sequence-length scalability. We would like to clarify that our method does not assume a fixed sequence length. All frames are jointly optimized in parallel without temporal windowing, and longer animations can be processed as long as GPU memory permits. We also preserve the original sequence lengths without truncation or interpolation during self-intersection repair for all the methods. To evaluate computational scalability, we benchmark 100 optimization rounds using an animation of a mesh with 5000 points and 10000 faces on a single NVIDIA RTX 5880 Ada GPU:

FramesRuntime (s)Peak Memory (GB)
90184.03.501
120237.54.663
150310.95.823
180371.96.867

Both runtime and peak memory increase approximately linearly with sequence length, demonstrating predictable scaling without sacrificing joint temporal optimization.

Regarding repair quality and motion fidelity, changing an animation's frame count also changes its temporal sampling and interpolated motion trajectory, making it difficult to isolate sequence length from changes in the underlying collision patterns. We therefore evaluate each animation at its original frame count, rather than artificially resampling sequences for quality comparisons.

Test set size and diversity. We have clarified the dataset composition and sequence-length statistics in the revised manuscript. Our evaluation includes 100 SMPL-H sequences (49 with 90 frames and 51 with 120 frames) and 30 non-SMPL animations from 24 distinct assets (15 sequences each from Objaverse and Truebone Zoo), covering diverse humanoid and animal characters with different geometries and skeletal structures. The non-SMPL sequences range from 17 to 177 frames, and all animations retain their original lengths without temporal resampling.

Thank you for the clarifications and additional experiments. The quantitative results on non-SMPL animations and the sequence-length scaling study largely address my concerns, and I am now more positive about the paper.

Thank you for taking the time to review our rebuttal and for your encouraging response! We are very glad that our new results have resolved your concerns and led to a more positive impression of our paper. If there are no remaining issues, we kindly wonder if you would consider updating your rating to reflect this more positive assessment. Please let us know if any further clarification would be helpful.

Reviewer_CburOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Proxyboros, a training-free framework for dynamic self-intersection repair in rigged mesh animations, considering the trade-off between collision resolution and motion fidelity. Its core component, Progressive Mesh-Rig Optimization (PMRO), alternates between mesh-space repair using the existing ISIR method and rig-space pose adaptation, progressively transferring local geometric corrections into skeletal motion through intermediate proxy meshes. Temporal Rig Parameterization (TRP) uses shared cubic B-spline controls to parameterize joint rotation corrections across frames and enforce temporal coherence. Additionally, the framework introduces surface-aware proxy guidance based on symmetric point-to-plane alignment with orientation-dependent weighting, along with pose regularization to constrain deviations from the original motion.

Critical strengths
  1. Overall competitive performance. As shown in Table 1, Proxyboros achieves generally better results than existing methods across most evaluation metrics. In particular, it demonstrates a substantial improvement in motion fidelity, especially in terms of Vertex Acceleration Difference (VAD).
  2. Progressive Mesh-Rig Optimization (PMRO). The proposed PMRO framework alternates between mesh-space repair and rig-space pose adaptation, explicitly considering the trade-off between collision resolution and motion fidelity. As shown in Table 2, introducing PMRO leads to significant improvements in pose preservation and motion fidelity compared to vanilla direct rig-space optimization.
  3. Temporal Rig Parameterization (TRP). TRP makes effective use of the inherent smoothness and local support properties of B-splines to impose temporal regularization through the parameterization of the optimization space. The ablation results in Table 2 demonstrate that TRP is one of the major contributors to motion fidelity, particularly in reducing VAD.
Critical weaknesses
  1. Unclear necessity of mesh-space proxies. While PMRO improves upon vanilla direct rig-space optimization (from Table 2), it remains unclear whether the intermediate mesh-space proxies are essential. A stronger baseline would directly optimize the collision loss in rig space while incorporating the same B-spline parameterization, pose regularization, and optimization budget as the proposed method. Such a comparison would help.
  2. SAG loss and loss balancing problem. The effectiveness of the proposed SAG loss remains unclear, as its orientation-dependent weighting also changes the overall loss magnitude. Moreover, Tables 1 and 2 suggest a clear trade-off between collision resolution and motion preservation, yet the choice of loss weights (λsag(\lambda_{\mathrm{sag}} and λpose)\lambda_{\mathrm{pose}}) is not sufficiently justified. An ablation study on these two loss weights would help clarify their influence on the trade-off and the effectiveness of the proposed SAG loss.
  3. Limited quantitative evaluation on heterogeneous rigs. Although generalizability across different mesh geometries and skeleton topologies is an important motivation of the paper, the main quantitative experiments are conducted on SMPL-H sequences. Evaluations on non-SMPL characters are primarily qualitative and supported by a user study. More systematic quantitative experiments across diverse rig structures would strengthen the generalizability claim.
Questions for authors
  1. Direct rig-space optimization baseline. Could the authors provide a comparison with direct rig-space collision optimization using the same B-spline parameterization, pose regularization, and optimization budget as Proxyboros? This would help clarify the necessity of intermediate mesh-space proxies.
  2. SAG loss and weight sensitivity. Could the authors provide an ablation study on λsag\lambda_{\mathrm{sag}} and λpose\lambda_{\mathrm{pose}} to investigate how their relative weights affect collision resolution and motion preservation? Additionally, a comparison with fixed or normalized SAG weights would help isolate the contribution of the orientation-dependent weighting mechanism.
  3. Quantitative evaluation on non-SMPL rigs. Could the authors provide additional quantitative results on non-SMPL characters, including repair quality and motion fidelity metrics across different mesh geometries and skeleton topologies, to further support the claimed generalizability?
  4. Post-hoc repair vs. collision-aware motion generation (Only for discussion). Since self-intersections often originate from motion generation or retargeting pipelines, could the authors discuss the practical advantages of post-hoc repair compared with incorporating collision constraints directly into these pipelines? In particular, could addressing collisions during motion synthesis offer greater flexibility to avoid the trade-off between collision resolution and fidelity to an already imperfect motion?
3 comments

We thank the reviewer for recognizing the significance and technical soundness of our work. Below, we address the main concerns.

Q1:

Performance over direct rig-space optimization. As discussed in the Introduction, directly optimizing collision objectives in rig space can induce excessive skeletal articulation, leading to large end-effector displacements and distortions in the original pose and motion. To further validate this issue, we introduce a direct rig-space optimization baseline that removes the progressive mesh–rig interaction while retaining the same collision objective, temporal parameterization, pose regularization, and matched optimization budget. The results are shown below.

MethodPVR (%) ↑IFR (%) ↑SRS (%) ↑VD (×10²) ↓VAD (×10³) ↓VAM (×10³) ↓
w/o PMRO95.0892.9990.3311.2321.3692.631
Full94.6489.7393.465.6300.6072.202

Direct rig-space optimization achieves higher collision reduction but substantially degrades pose preservation and motion fidelity, confirming the kinematic coupling issue illustrated in Figure 1 of the main paper. These results further demonstrate the importance of PMRO, which progressively incorporates small mesh-space repair updates into rig-space optimization, enabling effective collision resolution while preserving the original pose and motion. We have included the results and analysis in Appendix A.4 of our revised manuscript.

Q2:

We would like to clarify that collision repair and pose preservation inherently involve a trade-off, as resolving intersections requires pose adjustments that preservation metrics seek to minimize. Moreover, without a unique ground-truth repair, pose and motion deviation metrics serve as reference measures rather than absolute indicators of perceptual quality. Different deviation scores may correspond to equally plausible motions, with no universal threshold defining this distinction. We have included representative visual comparisons in the revised manuscript to illustrate such cases.

To further investigate loss balancing and the effectiveness of SAG, we conduct two additional experiments.

Loss-weight sensitivity. We fix λsag=1\lambda_{sag}=1 and vary λpose\lambda_{pose} from 10−310^{-3} to 10−110^{-1}.

metrics1e-33e-35e-31e-23e-25e-21e-1
PVR95.5495.0394.6493.7191.0588.6882.09
IFR91.2590.3789.7388.6186.2284.8181.85
VAD0.7270.6490.6070.5580.5040.4700.421
VAM2.2302.2102.2022.1892.1682.1592.146

Across a broad range of weights (1e-3 - 1e-2), our method maintains high repair quality (PVR: 93.71–95.54) and consistently outperforms PoseShield in motion fidelity (VAD: 0.558–0.727 vs. 3.071). As pose regularization increases, the repaired motion stays closer to the input, resulting in lower deviation metrics but progressively limiting collision resolution (PVR: 95.54 to 82.09). This behavior follows directly from the competing objectives of collision repair and pose preservation.

Effectiveness of Normal-Based Weighting. As analyzed in Appendix Table 5, SAG adaptively adjusts the local proxy guidance according to normal consistency, favoring more reliable repair directions.

Table 5 reference:

MethodsPVR (%) ↑IFR (%) ↑SRS (%) ↑VD ↓ (×10²)VAD ↓ (×10³)VAM ↓ (×10³)
Point2Point90.8386.2593.123.5790.4342.144
Point2Plane90.3985.7088.583.5460.4982.166
Lsag\mathcal{L}_{sag} w/o direction92.3087.8794.154.0650.4572.153
Lsag\mathcal{L}_{sag} w/o magnitude90.4085.6888.543.5600.5012.167
Full94.6489.7393.465.6300.6072.202

To distinguish this benefit from simply increasing alignment strength, we further evaluate the w/o magnitude variant with uniform weights of 2.0, 4.0, and 8.0, keeping pose regularization unchanged.

MethodsPVR (%)↑IFR (%)↑SRS (%)↑VD↓ (×10²)VAD↓ (×10³)VAM↓ (×10³)
LsagL_{sag} w/o magnitude90.4085.6888.543.5600.5012.167
-- w 2.092.0887.2090.334.0730.5252.177
-- w 4.093.7188.6291.914.8080.5592.190
-- w 8.094.7389.6893.546.3190.6962.235
Full94.6489.7393.465.6300.6072.202

Increasing uniform weights improves repair quality. Notably, even with a weight of 8.0, twice the maximum orientation-adaptive weight, the variant only achieves comparable repair quality and SRS to Full, while introducing larger pose deviations (VD: 6.319 vs. 5.630) and motion discrepancies (VAD: 0.696 vs. 0.607; VAM: 2.235 vs. 2.202). This demonstrates that orientation-adaptive weighting achieves a better balance between repair quality and motion preservation than uniform loss scaling.

Moreover, we clarify that VD should be evaluated jointly with repair quality, as smaller deviations may simply reflect incomplete intersection removal. Although our full method has a higher average VD than the ablated variants, Fig. 10 shows that substantially more samples achieve high PVR under the same VD constraints. In particular, increasing the VD threshold from 5.8 to 6.6 markedly increases the proportion of samples meeting stringent repair requirements. This suggests that the higher average VD is partly driven by challenging cases requiring larger pose corrections to resolve severe penetrations, rather than uniformly excessive deviations across all samples.

Q3:

Quantitative evaluation on non-SMPL rigs. We have added quantitative results on 30 non-SMPL animations from Objaverse and Truebone Zoo, covering diverse geometries and skeletal structures. As shown in Table 2 of the revised manuscript, our method achieves competitive repair quality with strong pose preservation and the best motion fidelity, further supporting its generalizability beyond SMPL-H.

MethodsIFR (%)↑SRS (%)↑VD↓ (×10²)VAD↓ (×10³)VAM↓ (×10³)
ISIR + skeleton fit81.4989.491.4634.97712.278
Mesh-Utg + skeleton fit87.3878.304.69317.52423.152
Ours84.5689.131.4311.72510.343

Q4:

Post-hoc repair vs. collision-aware motion generation. Incorporating collision constraints into motion generation is a promising direction, but existing approaches still struggle to fully eliminate self-intersections while preserving motion semantics and naturalness. Moreover, many of them rely on specific skeletal representations or training data, limiting their applicability to diverse rigged characters. In contrast, Proxyboros operates directly on existing rigged mesh animations without modifying upstream models or requiring additional training. This allows our method to handle motions from generation, motion capture, and retargeting, while also serving as a refinement stage to resolve residual penetrations in generated animations. Furthermore, Proxyboros can facilitate the construction of cleaner motion datasets to support future collision-aware generation. We have added a detailed discussion with relevant references in Appendix A.12.

Reviewer_SJoSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Proxyboros, an optimization framework for repairing self-intersections in rigged mesh animations while preserving the original motion. The method alternates short mesh-space repair steps with temporally coupled rig-space pose adaptation. Intermediate mesh proxies provide local repair guidance, which is incorporated through B-spline pose corrections, surface-aware alignment, and joint-position regularization. Experiments on SMPL-H sequences and diverse non-SMPL characters demonstrate a favorable balance between collision reduction and motion preservation.

Critical strengths
  1. The method addresses the tension between local geometric repair and coherent skeletal motion. Intermediate mesh proxies provide a well-motivated bridge between the two, without relying on a specific body model or learned motion prior.

  2. The results demonstrate a favorable balance between reducing self-intersections and preserving the original animation, with improvements in pose consistency and temporal fidelity.

  3. Component ablations, perceptual comparisons, and runtime analysis provide useful evidence for the proposed design. The evaluation also considers the trade-off between repair quality and motion preservation.

Critical weaknesses
  1. The ablation does not fully explain why normal-based weighting helps. Removing the weights also changes the overall strength of proxy alignment relative to pose preservation. This makes it difficult to tell whether the improvement comes from selecting more reliable repair directions or simply putting more emphasis on repair.

  2. The handling of brief, localized collisions needs more evidence. Shared spline controls encourage smooth corrections, but a collision lasting only a few frames may require a rapid, local adjustment. It remains unclear whether the method can resolve these cases without leaving residual intersections or unnecessarily changing nearby frames.

Questions for authors
  1. Does normal-based weighting retain its advantage when the overall alignment strength is matched across ablations? How does it handle valid repairs that require substantial changes in local surface orientation?

  2. How does the method perform on collisions lasting only a few frames, particularly regarding residual intersections and changes to neighboring collision-free frames? How sensitive is this behavior to the temporal control-point spacing?

4 comments

We thank the reviewer for the positive feedback on our well-motivated framework and its effectiveness in balancing repair quality and motion fidelity. Below, we respond to the main concerns.

Q1:

Effectiveness of Normal-Based Weighting. As analyzed in Appendix Table 5, SAG adaptively adjusts the local proxy guidance according to normal consistency, favoring more reliable repair directions.

Table 5 reference:

MethodsPVR (%) ↑IFR (%) ↑SRS (%) ↑VD ↓ (×10²)VAD ↓ (×10³)VAM ↓ (×10³)
Point2Point90.8386.2593.123.5790.4342.144
Point2Plane90.3985.7088.583.5460.4982.166
Lsag\mathcal{L}_{sag} w/o direction92.3087.8794.154.0650.4572.153
Lsag\mathcal{L}_{sag} w/o magnitude90.4085.6888.543.5600.5012.167
Full94.6489.7393.465.6300.6072.202

To distinguish this benefit from simply increasing alignment strength, we further evaluate the w/o magnitude variant with uniform weights of 2.0, 4.0, and 8.0, keeping pose regularization unchanged.

MethodsPVR (%)↑IFR (%)↑SRS (%)↑VD↓ (×10²)VAD↓ (×10³)VAM↓ (×10³)
LsagL_{sag} w/o magnitude90.4085.6888.543.5600.5012.167
-- w 2.092.0887.2090.334.0730.5252.177
-- w 4.093.7188.6291.914.8080.5592.190
-- w 8.094.7389.6893.546.3190.6962.235
Full94.6489.7393.465.6300.6072.202

Increasing uniform weights improves repair quality. Notably, even with a weight of 8.0, twice the maximum orientation-adaptive weight, the variant only achieves comparable repair quality and SRS to Full, while introducing larger pose deviations (VD: 6.319 vs. 5.630) and motion discrepancies (VAD: 0.696 vs. 0.607; VAM: 2.235 vs. 2.202). This demonstrates that orientation-adaptive weighting achieves a better balance between repair quality and motion preservation than uniform loss scaling.

Moreover, we clarify that VD should be evaluated jointly with repair quality, as smaller deviations may simply reflect incomplete intersection removal. Although our full method has a higher average VD than the ablated variants, Fig. 10 shows that substantially more samples achieve high PVR under the same VD constraints. In particular, increasing the VD threshold from 5.8 to 6.6 markedly increases the proportion of samples meeting stringent repair requirements. This suggests that the higher average VD is partly driven by challenging cases requiring larger pose corrections to resolve severe penetrations, rather than uniformly excessive deviations across all samples.

Large orientation changes. This concern appears to stem from a misunderstanding of SAG's weighting mechanism. SAG measures normal agreement between the current rigged mesh and its intermediate proxy, rather than between the input and final repaired surfaces. Since each proxy is generated through only a few mesh-space updates, large orientation discrepancies are less likely to arise within a single iteration. The progressive mesh–rig interaction then allows larger geometric corrections to be accumulated over successive iterations.

Q2:

Short-duration collisions. Our method can effectively resolves short-duration intersections while preserving the intended motion. We first clarify that our B-spline parameterizes residual pose corrections rather than the original motion, allowing localized repairs with temporally smooth adjustments. For collisions lasting only a few frames, restricting corrections to the intersecting frames could introduce abrupt pose changes and temporal jitter. Therefore, adjustments to neighboring collision-free frames are necessary to maintain motion continuity, while pose regularization limits unnecessary deviations. To further demonstrate this behavior, we have added a representative case with visualizations and frame-wise quantitative analysis in Appendix A.6. The results show that severe intersections concentrated in a few frames are successfully repaired, while the resulting vertex displacements vary smoothly across neighboring frames, preserving the original motion trajectory without abrupt corrections.

Temporal interval sensitivity. We further evaluate different B-spline control-point intervals (W=3,5,8,12,16), with results shown below.

MethodsPVR (%)↑IFR (%)↑SRS (%)↑VD↓ (×10²)VAD↓ (×10³)VAM↓ (×10³)
PoseShield92.0387.1188.615.8373.0712.646
w/o TRP94.5789.5994.205.5776.3005.444
Full (W=3)94.6289.7093.785.5901.0572.434
Full (W=5)94.6289.7393.595.6030.7642.251
Full (W=8)94.6489.7393.465.6300.6072.202
Full (W=12)94.5789.7593.375.6090.5362.180
Full (W=16)94.5889.7593.295.6480.5082.171

The results demonstrate that our method is robust to interval selection, with nearly identical repair quality and pose preservation across all configurations. Increasing the interval consistently improves temporal smoothness and reduces acceleration discrepancies, reflecting a controllable trade-off between local motion flexibility and temporal coherence. More importantly, all configurations consistently outperform PoseShield in motion fidelity, even at W=3W=3 (VAD: 1.057 vs. 3.071; VAM: 2.434 vs. 2.646). These results demonstrate that our method maintains its effectiveness across a wide range of temporal intervals.

Thank you for the clarifications and additional experiments. The weighting comparisons and temporal analyses largely address my concerns, and I am now more positive about the paper.

We appreciate your encouraging feedback and are pleased that our additional experiments and clarifications have helped address your concerns. Thank you again for your thoughtful comments, which have helped us improve the quality and clarity of our manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely appreciate the constructive feedback provided by all the reviewers and are pleased that Proxyboros has been recognized as well-motivated (SJoS, qxFY) and technically sound (Cbur). Its effectiveness (SJoS) and practical value (qxFY, tsGE) are also recognized. The main concerns focused on the necessity of progressive mesh-rig optimization (PMRO), the effectiveness of surface-aware guidance and temporal parameterization, and the robustness and generalizability of our evaluation.

In our detailed response below, we addressed these concerns with experimental results and analyses. All changes are highlighted in purple in the revised PDF.

The key revisions are summarized below:

  1. Non-SMPL quantitative evaluation (Sec. 4.2, Fig. 6, Tab. 2): We add quantitative comparisons on 30 heterogeneous rigged animations, demonstrating competitive repair quality and the best motion fidelity.
  2. Necessity of PMRO (App. A.4, Tab. 5): We introduce direct rig-space and one-shot optimization baselines to validate the necessity of intermediate proxies and progressive interaction.
  3. Effect of Surface-aware guidance (App. A.5, Tabs. 6–7): We add uniform-weight ablations and loss-weight sensitivity experiments to validate orientation-adaptive guidance.
  4. Temporal coherence and sensitivity (App. A.6, A.7, Fig. 11, Tab. 8): We analyze short-duration collision repair and B-spline interval sensitivity.
  5. Evaluation robustness and clarifications (App. A.2, A.8, A.12): We examine PVR sensitivity to voxel resolution and semantic partitioning, clarify within-part intersection evaluation, and discuss practical applications in motion generation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.