acceptodds
Under review as a conference paper at ICLR 2027

HyperCAPE: Learning Higher-Order Keypoint Structures for Category-Agnostic Pose Estimation

Abstract

Category-Agnostic Pose Estimation (CAPE) aims to localize semantic keypoints on unseen object categories. Recent CAPE methods increasingly exploit structural relations among keypoints, yet existing formulations predominantly rely on pairwise graphs, which do not explicitly represent group-wise structures jointly formed by multiple semantic keypoints. We propose HyperCape, a higher-order keypoint modeling framework that induces a latent hypergraph from contextualized keypoint representations. HyperCape sparsely assigns each keypoint to a subset of latent hyperedges, allowing related keypoints to form shared structural groups while permitting each keypoint to participate in multiple relations. The learned hypergraph not only enables higher-order feature interactions, but also directly guides coordinate refinement by modeling the relative geometry of keypoints within each hyperedge. Experiments on the MP-100 benchmark show that HyperCape outperforms existing graph-based CAPE methods, achieving gains of %p in mPCK and %p at [email protected] over the strongest prior method. Qualitative analyses further show that the learned hyperedges form coherent higher-order groups of structurally related keypoints across unseen categories. These results highlight the benefit of learned higher-order structure for accurate category-agnostic keypoint localization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Category-Agnostic Pose Estimation, Hypergraph Learning, Higher-Order Reasoning

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 3 reviews, above 85% of the others.
Reviewer_X5n2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the task of text‑driven category‑agnostic pose estimation. To address the limitation of prior graph‑based methods, which can only model pairwise relationships between keypoints and cannot explicitly capture high‑order group‑wise relations formed jointly by multiple keypoints, this paper introduces learnable latent hypergraphs for high‑order keypoint modeling. Unlike previous approaches that employ two separate structures for feature reasoning and geometric refinement, this work leverages the same hypergraph topology to jointly perform feature reasoning and spatial coordinate updates. Experiments are conducted on the MP‑100 benchmark dataset, and improvements over prior state‑of‑the‑art methods are achieved in terms of PCK and mPCK metrics.

Critical strengths

This method introduces learnable latent hypergraphs to address the limitation of pairwise graphs, which cannot explicitly model high‑order group‑wise relations among keypoints.

Critical weaknesses
  1. The experiments are insufficient: comparative evaluations are conducted on only one dataset, and the ablation studies are also relatively limited.
  2. The total number of hyperedges is fixed, without adequate explanation of its suitability for diverse inputs.
Questions for authors
  1. The performance is imbalanced across different object categories. Why does the mPCK drop on the insect category compared with the baseline?
  2. Will the latent‑hypergraph grouping fail when keypoints are occluded?
  3. Why is cross‑dataset validation absent in this work?
  4. The balancing loss only enforces utilization of hyperedges, but does not guarantee semantically meaningful group‑wise structures. Are there any quantitative metrics to evaluate the quality of learned hyperedge grouping?
  5. No runtime or GPU memory overhead is reported. What is the practical inference cost introduced by hypergraph aggregation and geometric refinement?
  6. The noise‑augmented gating module is only activated during training and disabled at inference time. How does this design affect robustness against OOD test samples?
3 comments
AuthorsOpenReview ↗

Response to reviewer X5n2 (1/3)

We sincerely thank the reviewer for the constructive feedback and for recognizing the contribution of learned latent hypergraphs to modeling higher-order group-wise relationships among keypoints. We have conducted additional experiments and analyses to address the reviewer's concerns and provide detailed responses below.

W1/Q3) Evaluation scope and additional ablation studies.

A) MP-100 is the established benchmark for category-agnostic pose estimation, and the major prior methods, including POMNet, CapeFormer, GraphCape, CapeX, CapeNext, and GenCape, follow its evaluation protocol without reporting separate cross-dataset transfer experiments. MP-100 itself comprises 100 object categories drawn from 13 source datasets, including COCO, AP-10K, CUB-200, and DeepFashion2, and evaluates generalization through five category-disjoint splits. Our evaluation therefore follows the established CAPE protocol and provides direct comparisons with prior methods under the same benchmark.

Although EdgeCape additionally reports results on AP-10K, it trains and evaluates within AP-10K rather than performing cross-dataset transfer from MP-100. Moreover, directly evaluating an MP-100-trained model on AP-10K risks train–test leakage, as AP-10K is one of MP-100's source datasets.

To further examine generalization across distinct object domains, we are conducting a cross-super-category evaluation, as discussed in our response to Reviewer sd3Q (W2). This experiment holds out entire super-categories during training, providing a stronger distribution shift than the standard category-disjoint splits. We will report the results during the discussion period and would be grateful if the reviewer could take a look.

To address the concern about limited experimental analysis, we have substantially expanded the revised manuscript with:

  • Learned graph comparisons (Sec. 4.3): Additional comparisons against text-conditioned adaptations of EdgeCape and GenCape, evaluating HyperCape against learned pairwise-graph methods rather than only fixed annotated graphs.
  • Hyperedge structure (Sec. 4.3 and Appendix B.3): Analyses of hyperedge cardinality and grouping quality, including comparisons with matched random assignments.
  • Hyperparameter sensitivity (Appendix B.2): Additional analysis of how increasing the hyperedge capacity affects the formation of higher-order groups.
  • Occlusion robustness (Appendix B.7): Quantitative evaluations under natural and synthetic occlusion, including hyperedge assignment stability.
  • Inference efficiency (Appendix C): Measurements of wall-clock latency, peak GPU memory, and parameter complexity relative to CapeX.

These additions provide a more comprehensive evaluation of HyperCape's structural modeling, robustness, and computational efficiency.

W2) Fixed number of hyperedges.

A) The fixed number MM specifies the maximum available hyperedge slots, not the number of groups that must be used for every input. Hyperedge memberships are predicted independently for each instance, allowing the effective structure to adapt to different keypoint configurations.

To verify this, we analyzed hyperedges containing at least two keypoints on MP-100 Split 1 test instances. With M=8M=8, the number of such hyperedges varies from 3 to 8 per instance, averaging 6.92 for categories with 9–15 keypoints and 8.00 for those with at least 25. With M=16M=16, it varies more widely (3–13 per instance, with averages of 6.51–12.07 across the same keypoint ranges). These results show that a fixed hyperedge capacity does not impose a fixed number of structural groups across diverse inputs.

Q1) Insect performance drop compared with CapeX

A) As shown in Appendix B.6 (Table 14), HyperCape outperforms CapeX in eight of nine object groups. The only exception is Insect, comprising fly and locust (2 of 100 categories; 6,000 of 300,000 queries), with a small decrease of −0.45−0.45 mPCK.

Both categories contain low-resolution grayscale images (160–192 pixels per side, versus the MP-100 median of 1,024) and densely packed keypoints. Of their 32–35 keypoints, 24 belong to six legs with highly repetitive textual descriptions. Our analysis shows that HyperCape substantially outperforms CapeX on fly body keypoints (38.1 vs. 30.0 mPCK), while the overall decrease is concentrated in the legs. For 84–95% of leg keypoints, predictions from both methods are closer to a ground-truth keypoint on another leg than to the correct one, indicating substantial ambiguity between similar legs. Since the learned hyperedges often group joints within individual legs, such confusion may also affect other joints through hyperedge aggregation, potentially contributing to the small performance decrease.

Manuscript revision. We have added a brief analysis of the Insect performance drop in Appendix B.6.

AuthorsOpenReview ↗

Response to reviewer X5n2 (2/3)

Q2) Robustness to occlusion.

A) In CAPE, occluded query keypoints remain as prediction targets and participate in structural reasoning, although MP-100 evaluates localization only on visible keypoints. In HyperCape, these keypoints remain in the hypergraph and can exchange information with other keypoints through their shared hyperedges.

Robustness to natural occlusion. We group MP-100 Split 1 test episodes by the proportion of visible query keypoints. HyperCape outperforms CapeX across all four visibility levels, with all 95% confidence intervals excluding zero. The largest improvement is observed under severe occlusion (+4.64 mPCK).

Occlusion levelVisible keypointsΔ mPCK (Ours − CapeX)
None100%+1.71
Mild75–<100%+1.92
Moderate50–<75%+1.69
Severe<50%+4.64

Robustness to synthetic occlusion. To directly assess the stability of learned hyperedge assignments, we additionally evaluate 7,500 MP-100 Split 1 episodes. We randomly select 30% of query keypoints and mask each with a square whose side length is 10% of the bounding box's maximum side length. We then compare hyperedge assignments before and after masking.

MetricResult
Masked keypoints with unchanged memberships95.3%
Instance-level assignment Jaccard similarity0.959
Occluded mPCK (HyperCape / CapeX)75.70 / 74.65

The learned hypergraph remains largely consistent under synthetic occlusion, with most masked keypoints retaining their original memberships. HyperCape also maintains higher absolute localization accuracy than CapeX after masking. Together, these results demonstrate that the learned grouping is largely stable under occlusion and continues to support accurate keypoint localization.

Manuscript revision. We have added Appendix B.7 with quantitative evaluations of localization performance under natural occlusion and hyperedge assignment stability under synthetic occlusion.

Q4) Quality of learned hyperedge grouping.

A) We agree that the balancing loss alone does not guarantee meaningful hyperedge grouping; it only encourages balanced utilization. However, hyperedge assignments are also learned end-to-end through the localization objective.

To evaluate grouping quality, we compare HyperCape against randomized assignments that preserve hyperedge cardinalities and the distribution of keypoint membership counts within each instance, controlling for hyperedge size and membership sparsity. We assess three properties on 400 test instances from 20 unseen MP-100 Split 1 categories:

MetricHyperCapeMatched random
Skeleton-connected hyperedges ↑33.2%11.2%
Relative spatial distance ↓0.781.00
Cross-instance membership Jaccard ↑0.8680.222

These results are consistent across three independent samples of 400 instances. Moreover, 51.6% of co-member keypoint pairs are separated by at least three hops on the ground-truth skeleton, showing that the learned groups are not restricted to directly connected skeletal neighbors. Together, these findings provide quantitative evidence of coherent structural grouping beyond balanced hyperedge utilization.

Manuscript revision. We have added a quantitative analysis of learned hyperedge grouping quality, including comparisons with matched random assignments, in Appendix B.3.

Q5) Runtime and GPU memory overhead.

A) We measured inference latency and peak GPU memory on the same 300 MP-100 test episodes, with three paired repetitions under identical conditions.

ModelTrainable Params.Peak GPU MemoryWall-clock Latency vs. CapeX
CapeX37.36M738.1 MBBaseline
HyperCape (M=8M=8)37.91M (+1.47%)732.8 MB+2.8% (median)

HyperCape incurs only 2.8% median wall-clock latency overhead, with comparable GPU memory usage. These results indicate that hypergraph aggregation and geometric refinement add minimal overhead to the overall inference pipeline.

Manuscript revision. We have extended the Parameter Complexity analysis in the Appendix C to include measured inference latency and peak GPU memory usage.

AuthorsOpenReview ↗

Response to reviewer X5n2 (3/3)

Q6) Inference-time noise and OOD robustness.

A) Noise-augmented gating serves as a training-time regularizer by exposing the model to perturbed hyperedge assignments. At inference, noise is disabled to ensure deterministic hyperedge assignments, while retaining the representations learned under these perturbations.

MP-100 evaluates generalization to unseen object categories through category-disjoint splits. To examine whether disabling noise affects this generalization, we evaluated both inference settings on Split 1. The results were nearly identical (79.44 mPCK in both settings, with differences below 0.03 percentage points at every PCK threshold). This indicates that disabling inference-time noise does not degrade performance on unseen categories in our evaluation, while ensuring deterministic predictions.

Reviewer_sd3QOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper suggests a new CAPE method that takes the recent directions of graph-based CAPE to the next level by integrating learned hypergraphs in order to share information between relevant groups of keypoints. They show this idea using a text-based CAPE model, based on CapeX, demonstrating SOTA performance for CAPE without requiring graph annotations.

Critical strengths
  • The paper is well-written and easy to follow.
  • The authors are willing to publish their code.
  • The idea of introducing hypergraphs into CAPE is reasonable, well-explained, and novel.
  • The paper demonstrates strong results, along with an extensive ablation study testing various aspects of the model.
Critical weaknesses
  • The model is text-based, making the most relevant comparisons CapeX and CapeNext, which both use a fixed annotated graph definition. This makes it difficult to determine whether the performance boost originates from using hypergraphs or from learning graphs dynamically. Several recent works have already demonstrated that learning dynamic graphs enhances performance (For example, EdgeCape/GenCape), so this finding is neither surprising nor novel. While I really like the idea of using Hypergraphs for CAPE, the current setting feels a bit unfair and fails to isolate the benefits of hypergraphs from other significant changes compared to other works. The work would be significantly stronger if it were demonstrated to be beneficial as an image-based CAPE model (which shouldn't require significant architectural changes) by surpassing predicted graph-based models in a similar setting. Or, alternatively, if it established a text-based baseline using EdgeCape or GenCape.

  • Missing super cross-category evaluation - prior works (such as EdgeCape and GenCape) demonstrated generalization capabilities by using a super-category experiment (established by Capeformer) which splits the MP-100 dataset into 8 supercategories, holding out each super-category in turn and training on the remaining ones. This evaluation is especially relevant for learned graph structures to show whether the model can generalize to unseen object categories. Demonstrating this capability would substantially strengthen the paper.

  • The model is very sensitive to hypergraph hyperparameters. According to Table 8, performance drops quite drastically when increasing the number of hyperedges beyond the selected value of 8.

Questions for authors

Please see weaknesses.

6 comments
AuthorsOpenReview ↗

Response to reviewer sd3Q

We sincerely thank the reviewer for recognizing the novelty of introducing learned hypergraphs into CAPE and for providing valuable feedback on our experimental evaluation. We have conducted additional analyses to address these concerns and provide detailed responses below.

W1) Comparison with learned graph-based methods.

A) We agree that comparisons against fixed annotated graphs alone are insufficient. To address this, we adapted EdgeCape and GenCape to the same text-conditioned CapeX pipeline (SwinV2-T backbone, frozen text encoder, and identical MP-100 Split 1 evaluation).

For Text-EdgeCape, we ported its graph predictor and Markov-biased attention, retaining its GT skeleton prior, adjacency supervision, and three-stage training procedure. For Text-GenCape, we incorporated its i-SVAE graph generator and graph fusion following the published method and hyperparameter settings, without annotated skeletons.

Method[email protected][email protected][email protected][email protected]mPCK
CapeX51.6577.9788.1692.6077.59
Text-EdgeCape52.1478.3888.5892.9678.02
Text-GenCape53.6778.7288.1792.3878.24
HyperCape55.4580.1589.1193.0779.44

Both learned pairwise-graph adaptations improve over CapeX, while HyperCape achieves the best performance across all PCK thresholds. These results show that HyperCape's gains cannot be explained solely by replacing fixed annotated graphs with learned structures.

Manuscript revision. We have added this comparison as Table 5, together with implementation details and quantitative analysis in Sec. 4.3 (Ablation and Analysis).

W2) Missing super cross-category evaluation.

A) We thank the reviewer for this suggestion. We are currently conducting cross-super-category evaluation following the protocol reported by POMNet and CapeFormer, which evaluates four held-out super-categories: Human Body, Human Face, Vehicle, and Furniture. For each evaluation, the held-out super-category is excluded from training. We will report the results during the discussion period and incorporate them into the revised manuscript.

W3) Sensitivity to hypergraph hyperparameters.

A) HyperCape is relatively stable around the selected configuration. As shown in Table 8 of the original submission (Table 9 in the revised manuscript), varying kk from 1 to 4 at M=8M=8 yields 78.97–79.44 mPCK, while M=6,k=3M=6, k=3 achieves 78.84. Performance decreases when the number of hyperedges is substantially smaller or larger than the selected value.

To better understand the degradation at larger MM, we analyze the learned hypergraphs with k=2k=2 fixed:

MMHyperedges with ≥3 keypoints / instanceKeypoints in ≥3-keypoint hyperedges
85.3192.9%
164.5770.2%
323.1550.9%

As MM increases, both the average number of hyperedges containing at least three keypoints and the proportion of keypoints participating in such groups decrease. These results suggest that excessively large MM disperses keypoints across smaller groups, potentially weakening higher-order structural interactions.

Manuscript revision. We have extended the hyperparameter sensitivity analysis in Appendix B.2 with these additional statistics on learned hyperedge cardinality and keypoint participation.

I thank the authors for their rebuttal.

I am inclined to raise my score, though I will wait for the missing super cross-category evaluation first.

AuthorsOpenReview ↗

Response to reviewer sd3Q

We thank the reviewer for the encouraging response. We have completed the requested cross-super-category evaluation.

Following POMNet and CapeFormer, we evaluate four held-out super-categories (Human Body, Human Face, Vehicle, and Furniture) using [email protected]. Since the original partition files were unavailable in the public repositories we examined, we reconstructed the partitions from MP-100 annotations and retrained all methods on identical splits with SwinV2-T. Following CapeX (Appendix A.3.2), we train for 50 epochs and fine-tune the gte-base-en-v1.5 text encoder for all text-conditioned methods.

Cross-super-category results ([email protected]). Bold indicates the best result among text-conditioned methods.

MethodBodyFaceVehicleFurnitureMean
GraphCape (Image)87.8187.1744.6756.1068.94
CapeX (Text)75.5764.1823.9136.8250.12
Text-GenCape (Text)79.1159.5933.6539.4152.94
HyperCape (Text)79.1572.0031.2949.2157.91

HyperCape achieves the highest mean [email protected] among text-conditioned methods, outperforming CapeX by 7.79%p and Text-GenCape by 4.97%p. While image-based GraphCape achieves higher accuracy, it benefits from annotated support images that provide explicit spatial cues. In contrast, text-conditioned methods must infer spatial relationships from descriptions such as "left" and "front," which can be ambiguous for unseen categories, as also discussed in CapeX. Nevertheless, HyperCape demonstrates stronger generalization than both fixed and learned pairwise-graph baselines under the same text-conditioned setting.

Manuscript revision. We have added the cross-super-category results and evaluation details to Appendix B.10 of the revised manuscript.

Reviewer_jni3OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HyperCape, a category-agnostic pose estimation framework that models higher-order structural relationships among semantic keypoints through a learned hypergraph. Unlike existing CAPE methods that primarily rely on pairwise keypoint graphs, HyperCape dynamically constructs a latent hypergraph from contextualized keypoint representations, enabling multiple keypoints to form shared structural groups. The proposed framework incorporates two key components: higher-order keypoint reasoning, which facilitates feature interactions through node-to-hyperedge-to-node aggregation, and geometry-aware keypoint refinement, which leverages the learned hypergraph to improve coordinate predictions. Experiments on the MP-100 benchmark demonstrate that HyperCape outperforms existing graph-based methods.

Critical strengths
  1. An interesting and well-motivated idea for structural modeling in CAPE. The paper introduces the concept of hypergraphs into CAPE, extending structural modeling beyond conventional pairwise keypoint relationships.
  2. The framework sounds like reasonable. It integrates higher-order keypoint reasoning with geometry-aware coordinate refinement, effectively leveraging the learned structural relationships for both feature interaction and keypoint localization.
  3. The method achieves improvements over existing CAPE approaches on MP-100. The ablation studies and qualitative visualizations further support the effectiveness of the proposed components.
  4. Well written and easy to understand.
Critical weaknesses
  1. Although the paper emphasizes higher-order structural modeling, the node-to-hyperedge-to-node aggregation in Equations (4)-(5) can be reformulated as message passing over a weighted pairwise adjacency matrix. And, the geometry-aware refinement in Equations (8) still relies on pairwise coordinate displacements. It is unclear whether the proposed framework captures higher-order dependencies or pairwise modeling.
  2. As described in Section 3.1 and Appendix A.2, the incidence matrix is constructed only once from the encoder and remain fixed through all decoder layers. But the keypoint representations are progressively refined during decoding. The fixed hypergraph may therefore prevent the model from correcting inaccurate structural assignments. It suggests to investigate whether dynamically updating hyperedge memberships during iterative refinement could further improve performance.
  3. The sparse hyperedge assignment mechanism does not guarantee the formation of genuine higher-order relationships. HyperCape assigns each keypoint to at most (k=2) of 8 latent hyperedges, while discarding memberships below the threshold. However, this mechanism only constrains the number of hyperedges assigned to each keypoint. Consequently, a hyperedge may contain only one or two keypoints, in which case the aggregation reduces to pairwise interactions rather than higher-order reasoning. Although Table 9 reports that all eight hyperedges are active under the default configuration, a hyperedge is considered active as long as it contains at least one keypoint, which does not establish that meaningful higher-order groups are consistently formed. The authors should report the hyperedge cardinality distribution and clarify whether a substantial proportion of hyperedges contain only one or two keypoints.
Questions for authors

See Weaknesses.

6 comments
AuthorsOpenReview ↗

Response to reviewer jni3 (1/2)

We sincerely thank the reviewer for the constructive feedback and for recognizing the motivation behind our higher-order structural modeling approach, the integration of feature reasoning and geometric refinement, and the improvements achieved on MP-100. We address the reviewer's concerns and questions in detail below.

W1) Higher-order vs. pairwise modeling.

A) We agree that the linear aggregation in Eqs. (4)–(5) can be reformulated as message passing over a weighted pairwise graph, and that Eq. (8) uses pairwise coordinate displacements. We do not claim that the aggregation alone provides greater expressive power than pairwise message passing. Rather, our goal is to explicitly model shared structural groups among multiple keypoints through a learned hypergraph and leverage these groups for geometric refinement.

Specifically, Eq. (8) conditions geometric interactions on the hyperedge representation eme_m and spatial spread ρm\rho_m, allowing the same keypoint pair to interact differently across hyperedges. Removing eme_m and γijm\gamma_{ijm} from Eq. (8), while retaining the learned hypergraph and geometric refinement mechanism, reduces performance across all PCK thresholds, as shown in the additional ablation results below. This supports the contribution of hyperedge-conditioned geometric refinement.

Manuscript revision. We have refined the description of hypergraph-based feature reasoning in the Abstract, Introduction, and Conclusion to avoid implying that linear hypergraph aggregation alone provides a strict higher-order expressivity advantage.

Additional ablation results (MP-100 Split 1).

Configuration[email protected][email protected][email protected][email protected]mPCK
HyperCape (fixed HH)55.4580.1589.1193.0779.44
w/o hyperedge-conditioned geometry54.5979.0188.1492.4978.56
Per-layer dynamic HH55.5580.2589.1292.8079.43

W2) Fixed vs. dynamic hyperedge assignments.

A) We keep hyperedge memberships fixed across decoder layers to provide a consistent structural reference, while allowing hyperedge features and geometry to evolve during iterative refinement. This avoids repeatedly reassigning structural groups as keypoint predictions change.

To examine whether dynamic reassignment improves performance, we evaluated a variant that recomputes HH at every decoder layer. As shown in the ablation table above, dynamic reassignment achieves nearly identical overall accuracy (79.43 vs. 79.44 mPCK), with small differences across individual PCK thresholds. We therefore retain the simpler design without repeated hypergraph construction.

AuthorsOpenReview ↗

Response to reviewer jni3 (2/2)

W3) Hyperedge cardinality and higher-order relations.

A) We intentionally do not enforce a minimum hyperedge cardinality, allowing the model to represent both pairwise and higher-order relations while leaving irrelevant hyperedges unused. This flexibility is particularly useful for CAPE, where keypoint configurations vary across categories.

To examine the learned structures, we analyzed the hyperedge cardinality distribution over the Split 1 test set. We found that 66.4% of hyperedges contain at least three keypoints, while 24.6% contain two, 7.6% contain one, and 1.4% are empty. These results show that HyperCape predominantly learns higher-order groups without explicitly enforcing a minimum cardinality.

Note that the 8/8 active hyperedges reported in Table 9 of the original submission (Table 10 in the revised manuscript) reflect utilization aggregated across test categories, whereas the cardinality distribution is computed per instance. Thus, all eight hyperedges can be utilized across the test set even though some remain empty in individual instances.

Manuscript revision. We have added the hyperedge cardinality distribution as Fig. 6, together with a quantitative analysis in Sec. 4.3 (Ablation and Analysis). We have also clarified the definition of active hyperedges in the caption of Table 10.

Thank you for the detailed response and additional experiments. Most of my concern in W2 and W3 has been resolved. To W1, I acknowledge that the additional ablations demonstrate the effectiveness of the proposed structural reasoning and geometry refinement modules. However, they do not fully prove that the improvements are attributable to higher-order interactions. Specifically, Eqs. (4)–(5) remain equivalent to weighted pairwise message passing, while the geometry refinement introduces hyperedge-conditioned contextual information that could potentially be incorporated into a suitably designed pairwise model. Could author briefly clarify what benefits the hypergraph formulation provides over pairwise modeling with comparable contextual information?

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive comments and thoughtful assessments of our work. We particularly appreciate their positive recognition of the motivation and novelty of introducing learned hypergraphs for higher-order structural modeling in category-agnostic pose estimation, as well as the effectiveness of our proposed framework.

We have carefully addressed the reviewers' concerns through additional experiments, quantitative analyses, and clarifications, and have revised the manuscript accordingly. We have provided detailed responses to all comments and questions in our individual replies to each reviewer. All revisions are highlighted in blue\textcolor{blue}{\mathbf{blue}} in the updated manuscript.

We hope that our responses and revisions address the remaining concerns and further clarify the contributions of HyperCape. We welcome any additional questions or feedback during the discussion period.

Dear Authors, Reviewers (jni3 and X5n2),

I would like to commend the authors for their detailed point-by-point responses and the comprehensive general response. I also thank Reviewer sd3Q for engaging in a very constructive and thorough discussion with the authors.

To Reviewers jni3 and X5n2: As we approach the end of the discussion period, the authors have provided specific responses to your questions. I kindly request that you take a moment to review their responses and leave a brief note on OpenReview. A simple confirmation of whether your concerns have been addressed and if your score remains the same is completely sufficient.

If I do not receive a follow-up before the discussion period closes, I will proceed with my evaluation based on the original reviews and the authors' written responses.

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.