acceptodds
Withdrawn from ICLR 2027

HIG-PCN: Higher Order Structure Guided Point Cloud Completion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Point cloud completion aims to recover complete object geometry from partial observations, supporting 3D shape understanding and robotic interaction when occlusion and limited viewpoints prevent complete scans. Existing methods commonly combine local geometric features with global shape context to infer missing surfaces. However, when an entire component is unobserved, nearby points may provide insufficient evidence, while a global representation may not explicitly preserve the relationships among the visible regions that constrain its geometry. For example, the remaining frame and bars of a partially observed chair back can jointly constrain the placement and arrangement of missing supports. Recovering such structures therefore requires considering how multiple regions relate to one another, rather than treating each missing region independently. This motivates explicit modeling of higher-order relationships that capture shared geometric context among sets of points. We propose HIG-PCN, a point cloud completion framework that represents these relationships using hypergraphs, where each hyperedge associates multiple related points. Its central idea is to combine local geometric groups with learned nonlocal groups and carry their shared context from shape encoding through point generation and refinement. In this way, evidence distributed across visible regions remains available when reconstructing unobserved geometry, without requiring semantic part annotations. Experiments assess completion accuracy and surface coverage on standard benchmarks, while controlled analyses show that both local and nonlocal group correspondences influence the reconstructed geometry.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hypergraph representation learning, Hypergraph neural networks, Point cloud completion

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_zS5COpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes HIG-PCN, a hypergraph-based framework for point cloud completion that aims to reconstruct missing geometric structures by modeling higher-order relationships among observed points. The authors argue that conventional methods often struggle with heavily occluded regions because local features provide insufficient information, while global representations may overlook important structural relationships. To address this issue, HIG-PCN introduces a dual-scale hypergraph encoder that combines local geometric hyperedges with learned nonlocal hyperedges. The resulting features are integrated into a multi-stage reconstruction pipeline. The authors conduct extensive experiments on various benchmarks, including PCN, MVP, and seen/unseen ShapeNet. The experimental results show that HIG-PCN achieves consistent improvements in Charfer Distance and F1-score over existing approaches.

Critical strengths
  1. The proposed method achieves consistent improvements over existing approaches on several point cloud completion benchmarks, particularly in unseen ShapeNet categories. These results suggest that the overall framework has some practical potential.

  2. The authors conduct component ablations and hypergraph membership perturbation experiments. Although these analyses do not fully establish the claimed benefits, they are useful attempts to evaluate whether the proposed representations influence completion performance.

Critical weaknesses
  1. The proposed framework largely combine existing techniques, including hypergraph feature aggregation, cross-attention, folding, and progressive refinement. Its methodological distinction from closely related work, particularly Hyper-PCN, is not sufficiently compelling. More importantly, the experimental results does not convincingly establish the necessity of the proposed higher-order structural guidance. In Table 4, dense refinement provides only marginal CD improvement while slightly degrading F1. Appendix G further shows nearly uniform decoder group assignments, raising concerns about whether meaningful group-specific information is actually used during reconstruction. The qualitative results in Figure 2 and the structural-region analysis in Appendix K also reveal substantial failures in recovering missing components, undermining the paper's central motivation.

  2. The experiments are not sufficient and rigor, and the quality of presentation is poor. Appendix D explicitly describes test set-based checkpoint selection on MVP, which raises concerns about the validity of the evaluation protocol. Moreover, the manuscript suffers from serious presentation issues. For example, mathematical expressions are awkwardly formatted. Figure 1 is overcrowded and difficult to follow, and several visualizations are too small to clearly demonstrate the claimed improvements. Overall, this paper requires substantial improvements in both experimental validation and technical presentation.

Questions for authors

1.Can the authors provide evaluation results using checkpoint selection independent of the test set?

2.Can the authors provide strictly controlled, end-to-end ablations that independently evaluate the contributions of local hyperedges, nonlocal hyperedges, and decoder-side group conditioning against parameter-matched alternatives? In particular, how do the authors explain the nearly uniform decoder group assignments reported in Appendix G?

3.Given the structural reconstruction failures shown in Appendix K, can the authors provide more comprehensive quantitative evidence that the proposed higher-order modeling actually improves the recovery of missing geometric structures, rather than merely improving aggregate Chamfer distance?

4.Can the authors provide a detailed comparison of parameter counts, FLOPs, inference latency, and memory consumption against closely related methods, particularly Hyper-PCN? Given the additional hypergraph processing, bidirectional attention, and refinement modules, it is unclear whether the relatively modest performance improvements justify the increased architectural complexity.

Reviewer_9aaBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HIG-PCN, a point cloud completion framework that employs higher-order relationships to improve the reconstruction of missing geometry from partial observations. The central idea is to preserve group-level geometric context throughout both feature encoding and point generation.

Specifically, the method introduces a dual-scale hypergraph encoder combining local geometric hyperedges constructed through spatial neighborhoods and nonlocal latent hyperedges learned through feature-based soft assignments. The resulting representations are incorporated into an anchor–group interaction module, where predicted coarse anchors exchange information with group tokens through bidirectional cross-attention. Subsequently, group-aware folding generates dense points, followed by progressive local–group refinement with bounded coordinate updates. An auxiliary intermediate-shape reconstruction loss is further introduced to improve the optimization of the folding and refinement stages.

The method is evaluated on PCN, MVP, and seen/unseen ShapeNet benchmarks, reporting competitive reconstruction accuracy in terms of Chamfer Distance and F-score. The paper additionally provides component ablations, incidence-permutation tests, checkpoint analyses, and structural-region diagnostics to investigate the contributions and limitations of the proposed design.

Critical strengths

(1) The paper addresses a meaningful challenge in point cloud completion, particularly the reconstruction of missing geometry when relevant contextual evidence is distributed across multiple observed regions.

(2) The overall architecture is technically coherent. The combination of dual-scale hypergraph encoding, anchor–group interaction, group-aware folding, and progressive refinement provides a systematic mechanism for incorporating group-level context into different stages of point generation.

(3) The method reports competitive results across multiple established benchmarks, including PCN, MVP, and seen/unseen ShapeNet categories.

(4) The paper provides relatively extensive diagnostic analyses, including incidence-permutation tests, intermediate-shape supervision studies, and structural-region evaluations. The explicit discussion of negative results and methodological limitations is appreciated.

Critical weaknesses

(1) The main contribution appears to be an architectural combination of established mechanisms rather than a fundamentally new higher-order geometric reasoning approach. In particular, Hyper-PCN (CVPR 2026) has already introduced hypergraph-based high-order correlation modeling for point cloud completion, incorporating hypergraph feature propagation, anchor-based reasoning, and subsequent point generation. Although HIG-PCN differs in its dual-scale incidence construction and explicit propagation of group context through the decoder, the proposed components largely rely on established techniques, including FPS/kNN grouping, prototype-based soft assignment, bidirectional cross-attention, folding-based decoding, and residual refinement.

More importantly, the paper does not sufficiently demonstrate that its specific higher-order representation offers advantages that cannot be achieved through simpler alternatives, such as weighted graph aggregation, conventional latent tokens, or standard cross-attention. The novelty and necessity of the proposed combination therefore remain insufficiently established.

(2) A particularly concerning observation is reported in Appendix G.2. The decoder-side group assignments exhibit nearly maximum normalized entropy (0.9992), an effective group count of 16.0, and almost identical spatial centers. Furthermore, perturbing the decoder assignments produces negligible changes in reconstruction metrics, while the encoder–decoder soft NMI is only 0.00010.

These observations suggest that the decoder assignments may have degenerated into nearly uniform memberships rather than learning meaningful and distinguishable geometric groups. Although the decoder can still benefit from global context and subsequent attention operations, these results weaken the central claim that informative higher-order group relationships are explicitly preserved and exploited throughout reconstruction.

The paper should clarify what information remains specific to individual decoder groups and whether the proposed group-conditioned decoder offers benefits beyond standard global feature aggregation or learned latent-token conditioning.

(3) The current ablation studies do not sufficiently isolate the contributions of the dual-scale hypergraph encoder, anchor–group interaction, and group-conditioned generation. In Table 4, adding group-token dense refinement changes MVP CD-L2 from 7.31 to 7.29, while F1 decreases from 0.409 to 0.407. Appendix F.2 further reports that the CD improvement is not statistically significant at the category level (p = 0.117). The subsequent improvement obtained through intermediate-shape supervision does not necessarily demonstrate the effectiveness of higher-order group modeling, as auxiliary supervision can also benefit conventional coarse-to-fine architectures.

Moreover, the incidence-permutation experiments primarily establish that the trained network depends on the correspondence between points and hyperedge memberships. Since the perturbations disrupt the learned feature–geometry alignment, the resulting performance degradation does not demonstrate that higher-order modeling is superior to ordinary graph aggregation or attention-based context modeling.

More comprehensive, matched-training ablations are needed, including local-only versus latent-only versus dual-scale hypergraphs, group tokens versus conventional latent tokens, and refinement with versus without group-specific context.

(4) The paper motivates higher-order modeling through examples involving completely missing structural components, such as chair-back supports. However, the structural-region diagnostic in Appendix K reports substantially lower missing-region recall for HIG-PCN than the reference methods in both examined regions. For example, the chair-backrest recall is 0.479 for HIG-PCN versus 0.705 for SymmCompletion at the reported threshold. Although these two examples do not establish a general performance ranking, they do not support the claimed advantage in recovering missing structures. A larger, systematically selected evaluation of genuinely unobserved geometric regions is necessary.

In addition, the relationship between the main MVP result (CD-L2 of 4.68 in Table 3) and the substantially different results reported in the checkpoint and mechanism analyses (e.g., Table 16) is not sufficiently clear. Appendix D also acknowledges test-set-based checkpoint selection for MVP. Although the authors provide fixed-epoch analyses, these do not fully resolve concerns regarding model selection for the primary benchmark result. The paper would benefit from an explicit mapping between experimental configurations, checkpoint-selection procedures, and reported results, together with selection-independent comparisons and repeated training runs.

(5) Several tightly related works are missed in the related works [1-3]. [1] Li, W., Yang, Z., Han, W., Man, H., Wang, X., & Fan, X. (2025, April). Hyperbolic-constraint point cloud reconstruction from single rgb-d images. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 39, No. 5, pp. 4959-4967). [2] Fu, Z., Zhang, J., Wang, L., Xu, L., Laga, H., Guo, Y., ... & Bennamoun, M. (2025). CompletionMamba: Taming state space model for point cloud completion. IEEE transactions on image processing. [3] Yu, X., Liu, Z., Wang, Y., Zhou, J., & Lu, J. (2026). ProtoComp++: Diverse Point Cloud Completion with Controllable Prototype. IEEE Transactions on Pattern Analysis and Machine Intelligence.

Questions for authors

(1) Compared with Hyper-PCN (CVPR 2026) and conventional latent-token attention mechanisms, what specific capability does the proposed higher-order group representation provide that cannot be achieved through standard feature aggregation? Could the authors provide parameter- and training-matched comparisons against these alternatives?

(2) How do the authors explain the nearly uniform decoder-side group assignments reported in Appendix G.2? If these groups are spatially indistinguishable and their membership perturbations have negligible effects, what evidence supports the effectiveness of group-guided reconstruction rather than generic context-conditioned decoding?

(3) Can the authors provide systematic evaluations of completely unobserved geometric components, rather than relying primarily on overall CD and F-score? In particular, how does HIG-PCN compare with Hyper-PCN and other baselines in terms of missing-region recall and structural consistency under matched evaluation protocols?

(4) Could the authors clarify the exact relationship between the model configurations underlying Table 3, Table 4, and Appendix H? Which configurations and checkpoint-selection procedures produce the reported benchmark results, and do the principal conclusions remain valid under selection-independent evaluation and multiple training seeds?

Reviewer_NoxnOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces HIG-PCN, which combines local geometric hyperedges with learned nonlocal groups for point cloud completion through anchor–group interaction, group-conditioned generation, and progressive refinement. The paper reports improvements on PCN, MVP, and seen/unseen ShapeNet benchmarks, together with incidence-permutation experiments, component studies, and checkpoint-selection analyses.

Critical strengths

The idea of using soft latent-group incidence based on the hypergraph representation sounds reasonable.

Critical weaknesses
  1. Poor representation. The presentation of the algorithms, the mathmatic definition, the figures are very difficult to follow. Many symbols are not clearly defined and explained, eg. CC in Eq 1, rˉ\bar{r} and fˉ\bar{f} n Eq. 2, FF and WW in Eq. 3, q^j\hat{q}_j in Eq. 7, and so on.
  2. Untrustworthy data. The average CD values (6.18 and 0.53) of the proposed method in Table 1 and Table 2 are not consistent with the per-category CD. Furthermore, there is no comparison with the method proposed by Li et al. (2026b), which proposed the hypergraph idea for point cloud completion.
  3. Unclear ablation study. It is not clear what dataset is the ablation study conducted on. The reported F1 value is much lower than the reported F1 values in Table 1\2\3.
  4. Unclear correspondence with the documented experimental configurations. Table 3 reports MVP CD/F1 of 4.68/0.568, whereas Appendix Table 16 reports 7.44/0.452 for test-selected HIG-PCN and 8.30/0.456 at epoch 420 under the same stated dataset and metric scale. These values may come from different configurations, but the manuscript does not clearly identify the model, supervision settings, and checkpoint underlying the headline result or its relationship to the audited models. This limits a reliable assessment of the claimed accuracy improvement.
  5. The independent benefit of the central grouping mechanism remains insufficiently established. The dual-scale incidence-permutation experiments provide evidence that the audited model depends on both encoder-side relations. However, Appendix G.2 and Table 14 show nearly uniform decoder memberships and little effect from permuting them, leaving the distinctive benefit of group-conditioned decoding over broadly pooled context unclear. The existing refinement ablations also do not replace matched training comparisons isolating the hypergraph branches and anchor–group interaction. The evidence therefore remains insufficient to establish the specific advantage of preserving differentiated group structure throughout generation.
Questions for authors
  1. Which configurations, supervision settings, and checkpoints produced the main-table and appendix results? Do the decoder-group diagnostics apply to the model used in the main table? Please also explain how Table 2’s seen ShapeNet-34 average of 0.53 is obtained from 0.31/0.45/0.96.
  2. What are the effects of separately removing or replacing the local hypergraph branch, nonlocal grouping branch, and anchor–group interaction while keeping positional encoding, loss settings, and training budget unchanged?

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

Related papers

Discussion (0)

Sign in to comment.