acceptodds
Withdrawn from ICLR 2027

Learning 3D Attachment Rankings from Geometric Proxies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Can geometric supervision teach a model which component an assembled part attaches to? We study parent ranking from a reference-pose mesh with supplied component assignments. C2LT learns from estimated patch-connection labels computed from ShapeNet geometry, without annotated parent relations, and scores pairs of surface patches to rank candidate parents. We evaluate the trained model without fine-tuning on 4,069 queries from 931 PartNet-Mobility objects. The full model reaches 44.74% parent accuracy, compared with 38.49% for a model trained under the same protocol but without cross-patch attention. Its advantage remains when every candidate receives the same number of patch-pair scores. However, when sampling guarantees at least one patch per component and allocates the remaining patches by area, the full model reaches 65.37%, whereas choosing the candidate with the largest whole-component surface area in the supplied mesh reaches 90.17%. Cross-patch attention therefore improves the learned predictor, but this improvement does not surpass a simple geometric rule on the evaluated task.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
3D representation learning, geometric supervision, attachment prediction, part connectivity, structured prediction, 3D geometry

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_1c2COpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The proposes a method to predict the parenthood relationships between parts of a 3D object. It does so through a transformer architecture that combines both local patch features with object-level context, training on a synthetically generated dataset in which relationships have been inferred by geometric positions.

Critical strengths

Frankly, it is hard for me to identify a strength of this paper. If pressed, I would say that the problem it studies is important, and that I do believe that predicting object parenthood is a worthwhile effort that could have a myriad of applications if done correctly.

Critical weaknesses

One critical weakness of this manuscript is the writing, which is consistently vague, repetitive and bordering on meaningless. At times it seems to reference missing context, as if there were parts of the text that were edited out but their references remain. Almost every paragraph reads as unnecessarily defensive, comparing every introduced concepts to separate, irrelevant concepts. Terminological choices are non-standard and often confusing. Here are some examples from the first pages, though they are by no means exhaustive and this writing issue persists through the entire text:

  • L30 "not merely a nearby handle" what is the meaning of this aside? is this referring to the door handle? the door handle is obviously connected by parenthood to the door, though in the reverse direction.
  • L34 "rather than surface reconstruction" why bring up surface reconstruction at all? this seems completely irrelevant to the task at hand.
  • L46-47 "ShapeNet patches supply estimated connection labels without parent annotations" this sentence cannot be understood at the point of the paper at which it appears. What does "supply" mean in this sentence? Are they part of ShapeNet? Inferred from ShapeNet? Learned?
  • L51 "The ShapeNet CAD meshes used for training" this is the first time these meshes are introduced.
  • L58 the method's name is never introduced before the acronym is used
  • L88 "address related geometry and structure problems" this is a very vague sentence. Isn't the point of this section to tell the reader how they are related?
  • L97 "both use assembly supervision and output joint/mate information [...] accuracies are not directly comparable". This makes sense, but it seems like joint/mate relationships can lead to parenthood relationships, so why not attempt to extract them from it? More critically, if assembly datasets are available and can be used to learn very similar relationships, why not use those for this method instead of inferring relationships geometrically?
  • L110 "A surface patch (a chart in the implementation)" why is this referencing the implementation? why the need for two separate names?
  • L115 "A physical component is a supplied rigid link" Supplied by whom?
  • L116 "ShapeNet validation selects checkpoints" what does this sentence mean? Selects how?

I'll stop enumerating here, as this poor writing makes it quite hard to understand the contributions of the following sections if there are any. The steps of the method are not clearly stated at any point, and notation and nomenclature regularly appear without or with minimal explanation (e.g., "geometry derived center", "nominal radius", what do these things mean?)

While the above is reason enough to reject a paper (indeed to desk-reject it, in my humble opinion), I also believe the paper contains a fundamental theoretical weakness: namely, it derives attachment rankings geometrically using simple geometric heuristics and then trains a model to predict these rankings instead of using the geometric heuristics directly (a "cannonball to kill a fly" approach). I realize Section 5.4 intends to address this, unfortunately, the results are not conclusive and the section is mostly unreadable (see examples above).

I do not know if any human was involved in the production of this text or these results; if they were, I would encourage them to keep working on this research problem, but perhaps with a different approach entirely.

Questions for authors

I don't have questions for the authors.

Reviewer_i9eKOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies attachment-parent ranking on assembled 3D meshes and proposes C2LT/Grouped, which learns local patch features, uses cross-patch Transformer context, scores patch pairs, and aggregates them into component rankings using supplied component assignments. Trained with geometry-derived proxy labels from ShapeNet and evaluated on PartNet-Mobility, it shows that cross-patch attention improves accuracy from 38.49% to 44.74% (+6.25 points), and by +4.10 points under equal pair opportunities, but the learned ranker still loses to simple geometric rules such as choosing the largest component by surface area (62.25% original, 90.17% with full coverage). The authors conclude that cross-patch attention has a real architectural effect, but proxy-supervised learning does not yet establish a task-level advantage over strong input-derived geometric cues, and that coverage, pair opportunities, and component-ID shortcuts must be reported.

Critical strengths

1.A well-controlled architecture comparison that isolates cross-patch information exchange. The paper’s strongest scientific contribution is the matched comparison between the full Grouped model and a no-cross-patch-attention control that shares the tokenizer, data, training budget, initialization, Transformer depth, and input cues. The authors further control for unequal patch-pair opportunities by equalizing the number of scored pairs per candidate. This is consequential because the central claim—that object-wide attention improves parent ranking—would otherwise be vulnerable to trivial explanations such as more parameters, different training, or candidates simply having more chances to obtain a high maximum score. The fact that the gain persists under these controls makes the architectural finding credible even though the learned model does not beat the strongest geometric rule.

2.An unusually honest and informative evaluation against strong geometric baselines and benchmark shortcuts. The paper does not stop at showing that Grouped beats its no-attention control. It directly compares against dense geometric support, nearest-neighbor geometry, full-proxy recomputation, patch-count rules, largest-component-surface-area rules, and component-ID rules, and it analyzes coverage, candidate availability, ID permutation, geometric-group permutation, and object/category macro-averaging. This is consequential because it reveals that the apparent learned ranking is still below simple input-derived cues and that the benchmark contains strong coverage and ID-order shortcuts. This kind of analysis substantially increases the paper’s methodological value and supports acceptance as a careful, credible study—even if its headline result is partly negative.

Critical weaknesses
  1. The proposed learned ranker does not establish a task-level advantage over simple input-derived rules, which undermines the paper’s methodological contribution. The paper’s own results show that Grouped reaches 44.74% on the original patches, while the largest-component-surface-area rule reaches 62.25%; under full component coverage, Grouped reaches 65.37% (area-covered) and 44.86% (uniform-covered), while the surface-area rule reaches 90.17% in both settings. The equal-pair experiments show that cross-patch attention helps relative to a no-attention control, but this is an internal architectural effect, not evidence that the learned model is a better solution to the task. This is consequential because the paper is framed as introducing a geometry-supervised predictor for attachment-parent ranking; if that predictor is consistently beaten by a trivial geometric rule, the practical and scientific value of the proposed method is limited. A diagnostic study of why the model fails could still be valuable, but the current paper does not fully reposition itself as such.

  2. The evaluation is not an independent confirmation, and the benchmark contains strong shortcuts that weaken the generality of the central comparisons. The authors explicitly state that the PartNet-Mobility target set has been examined repeatedly, that follow-up comparisons are exploratory, and that no independent target validation was performed. In addition, the largest-component-ID rule reaches 62.60% but falls to 20.38% after ID permutation, showing a strong dataset-order shortcut; patch-count and surface-area rules are also strongly affected by sampling coverage and candidate availability. The proxy labels are heuristic geometric quantities, not physical contact or human-validated attachment labels. This is consequential because the paper’s conclusions depend on comparisons on this fixed, repeatedly inspected cohort, and the presence of strong coverage, ID-order, and sampling biases makes it difficult to know whether the reported architectural gain would transfer to a clean, independent benchmark. Without such confirmation, the paper’s negative task-level conclusion is plausible but not decisively established as a general property of the method.

Questions for authors
  1. In the uniform-covered setting, the largest-surface-area rule reaches 90.17% while Grouped reaches 44.86%. In what concrete use case do the authors believe the learned ranker still adds practical value? Why it matters: The answer would determine whether I view the paper primarily as a negative/diagnostic study or as a method contribution. If the model is only useful when whole-component area is unavailable or unreliable, the paper should state that scope explicitly and support it with targeted results.

  2. Are there any pre-registered or independent evaluation results on a PartNet-Mobility subset or another dataset that was not repeatedly inspected during analysis? Why it matters: The authors acknowledge that the target cohort has been examined repeatedly and that follow-up comparisons are exploratory. Independent confirmation would materially affect my soundness assessment and the generality of the negative task-level conclusion.

As for future works,

  1. Evaluate whether Grouped and the surface-area rule make complementary errors, especially on queries where the area rule fails. The paper reports some fusion/tie-break results, but a focused analysis of area-rule failure cases would clarify whether learned scores contain any task-relevant signal beyond the strongest geometric prior.

  2. Construct a shortcut-controlled benchmark that removes or explicitly controls component-ID ordering, candidate coverage, and area-correlated sampling, then compare Grouped against matched geometric rules. This would directly test whether the observed cross-patch attention gain reflects attachment structure rather than benchmark-specific cues.

Reviewer_Gkg3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is hard to follow due to its poor AI-written style! The following is my summary of what I understand so far.

The paper studies whether one type of assembly relationship, “attached to” can be inferred from a multi-component 3D model using geometric information only. It designs a set of heuristic geometric proxies on ordered pairs of local patches, namely surface overlap, boundary proximity, normal agreement, and occupancy agreement, and uses them to form compatibility, collision-proxy, and invalidity-proxy scores (Cij,yijcoll,yijinv)(C_{ij}, y_{ij}^{coll}, y_{ij}^{inv}), which are further combined into a pair score. A component-level ranking score S(u,v)S(u,v) is then obtained by taking the maximum score over all patch pairs between two components. The key hypothesis is that these geometry-derived proxy relations contain useful information that can transfer to the ranking of attachment parents.

To learn and transfer these proxy relations and compute S(u,v)S(u,v) in a more efficient way, a Transformer is trained to produce contextualized patch features, and an MLP is trained to predict (Cij,yijcoll,yijinv)(C_{ij}, y_{ij}^{coll}, y_{ij}^{inv}) for patch pairs. S(u,v)S(u,v) is then computed from these predictions. The networks are trained on ShapeNet data without attachment-parent annotations and tested on PartNet-Mobility data with ground-truth parent labels.

Under the original sampling protocol, the learned model achieves 44.74% accuracy. When every component is guaranteed to be sampled and the remaining patches are allocated by component area, the learned model reaches 65.37% accuracy, while a simple criterion that selects the candidate component with the largest surface area achieves 90.17% accuracy. These honest but contrasting findings leave the practical value of the learned model unclear.

Critical strengths

I find the strengths of the proposed method limited. Nevertheless, the paper deserves credit for reporting extensive diagnostic experiments and negative results, including the substantially stronger performance of a simple component-area criterion.

Critical weaknesses

My analysis below focuses on the technical and writing aspects.


Technical aspects

  1. Motivation

In principle, the work needs to provide a deeper discussion of why geometry alone could provide sufficient clues for deriving attachment relationships. In reality, attachment is determined by motion and by how two parts are physically linked, for instance, via a hinge. Therefore, the relation could be represented as a tuple such as ((partA,hinge,partB)((part_A, hinge, part_B). Furthermore, semantic information could help utilize common knowledge for inferring assembly relationships. Unfortunately, PartNet does not provide all of the above information for learning and practical use. Seeking a real assembly-oriented 3D collection with richer labels may therefore be necessary to pursue this research; there is no free lunch in deriving such information from part geometry alone.

Furthermore, I do not understand the practical motivation of this work: in what kinds of scenarios would one need to infer attachment relationships for a multi-component object? I feel that the research question is somewhat artificial.

  1. The construction of patches and groups.

First, I am not sure why the groups are constructed as follows: “A geometric group collects connected faces, splits large regions by face-centroid clustering, and merges small fragments.” A given input mesh may have multiple connected components, so why not use each connected component as a group directly? Furthermore, why is the splitting necessary, and for what purpose?

Second, I find that a patch is sampled over the object surface and can therefore cover multiple groups/components. The current implementation uses the anchor face to obtain its group or component ID. I do not think this is a good design. All points within a patch should ideally lie on the same mesh component. Crossing component boundaries may introduce irrelevant information and make the patch features confusing for later use. Ideally, I would suggest the following relation:

patch⊂group⊂component\text{patch} \subset \text{group} \subset \text{component},

where the patch area can be uniformly defined and the patch-sequence length can vary.

  1. Group ID encoding

The value of the group ID is embedded as a feature in Eq. (1). I consider this a questionable design because it introduces an artificial prior into the network, as also suggested by the ablation study. A better design may be to randomize the group IDs during training and use a relational encoding that depends only on whether or how two groups are related, rather than on their absolute numeric IDs.

  1. The proxy-based ranking score.

The experiments in Section 5.4 do not support the effectiveness of the proxy-based S(u,v)S(u,v) for ranking. Even when using the “direct full-proxy” rule, the accuracy is only 46.72%. This suggests that the limitation does not simply come from the neural network's approximation of the proxies; rather, the proxy itself appears to provide a weak signal for parent ranking. This experiment therefore directly challenges the central research hypothesis of the paper.

  1. Transformer design.

Since the original proxy values are computed between two patches, it is unclear why a Transformer needs to perform attention over all patches of an object. What additional information does this object-wide context provide for learning the patch-pair relation? I do not find sufficient discussion motivating this design. There is a “Local” baseline, but its model capacity is clearly smaller than that of the full pipeline, which additionally contains the Transformer blocks. Therefore, this comparison does not cleanly isolate the benefit of object-level context.

  1. Spatial baseline.

It is unclear why the Spatial baseline is not considered the default choice, since it achieves better performance than the full Grouped model in Table 9.

  1. Patch vs. component.

Since the test examples already provide physical components, why not treat each component as the basic unit and derive suitable geometric proxies directly for component-level ranking? The current patch-to-group-to-component construction seems unnecessarily complicated for a task whose final output is defined between components.


Writing aspects

  1. I find the paper unfriendly to readers. It is difficult to understand what key findings the authors would like to communicate. The paper reads more like a poorly organized AI-drafted tchnical report than a paper with clear emphasis and a coherent scientific story. In particular, the key motivation for designing these geometric proxies, and why these proxies should be expected to work for attachment inference, are never clearly communicated. The last sentence of the abstract, “this improvement does not surpass a simple geometric rule on the evaluated task,” also immediately raises the question of whether the proposed method provides sufficient value for publication.

  2. There are too many terms that are not clearly defined or sufficiently discussed. Here I list only a few examples from the first few pages. a. Undefined or insufficiently introduced terms: C2LT; patch connection. b. Unclear definitions/details:

    • Line 160: rotation RiR_i. How is it computed, e.g., via PCA?
    • Line 162: XilocX_i^{loc}.
    • Eq. (3): the heuristic choices require better justification. Visual comparisons would also help explain what these proxy quantities actually capture.
    • Without fine-tuning. I find this statement confusing, I don't see how fintuning can be imposed to the presented work, as there are no parent-ranking labels for use.
Questions for authors

No additional questions, see "Critical Weaknesses" for my main feedback.

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

Related papers

Discussion (0)

Sign in to comment.