acceptodds
Desk rejected at ICLR 2027

PartMosaic3D: COUPLING PART-AWARE DECOMPOSITION AND JOINT CONSTITUENT SELECTION FOR 3D OBJECT MOSAICS

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

We study 3D object mosaic synthesis from a fixed gallery: reconstructing a textured target mesh by placing whole objects retrieved from a gallery of 3D assets. Any solution makes two coupled decisions—how the target is partitioned into regions, and which object occupies each region—that existing approaches make implicitly or greedily. In this work, we present PARTMOSAIC3D, a formulation that makes both decisions explicit: regions are nodes of a learned semantic part hierarchy, refined where retrieval under a frozen shape representation is poor, and constituents are chosen jointly over the region adjacency graph with fidelity, fit, reuse and interpenetration costs. The fidelity-and-reuse core of this problem is an exactly solvable min-cost flow that contains the baseline’s greedy reuse heuristic, and the full problem is solved to certified optimality within candidate pools. A decomposition × selection factorial with matched-granularity controls tests whether and when each decision matters. On four targets, parts with joint selection raise coverage from 0.527–0.703 to 0.725–0.918 over voxels with greedy selection at matched granularity, and at the baseline’s granularity parts reduce the protrusion error from 1.71% to 0.30% and from 1.38% to 0.32% where parts are contiguous. Joint selection helps through its fit and overlap terms, cutting interpenetration volume by 30–54% on three targets, whereas exact optimization of fidelity and reuse alone changes the mosaics only marginally; retrievability-driven refinement showed no benefit over size-driven subdivision at the thresholds tested, and amodal completion did not help.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Object Mosaics, Part-Aware Decomposition, 3D Shape Analysis, Global Assignment, 3D Object Retrieval, Computational Creativity, 3D Geometry, Non-Photorealistic Rendering

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_QXAjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PartMosaic3D constructs 3D object mosaics by decomposing a target mesh into regions and selecting whole objects from a fixed gallery. It combines PartField-based decomposition with joint selection using similarity, geometric fit, reuse, and overlap costs.

Critical strengths
  1. The formulation clearly separates decomposition and selection, and the matched-region-count comparisons provide useful controls.
  2. The paper reports informative ablations and transparently discusses negative results, including the limited benefit of retrieval-driven refinement.
Critical weaknesses
  1. The comparison with uniform voxels does not clearly establish the added value of semantic part decomposition. PartField may perform better simply because its regions follow the object’s geometry more closely. Matching the number of regions does not control their shapes or connectivity. A comparison with a geometry-based segmentation method that does not use semantic features would help clarify whether the gains come from semantic information or simply from better-aligned region boundaries.
  2. The manuscript includes a pipeline diagram but no rendered examples or visual comparisons of the final mosaics. For a visual synthesis task, geometric metrics alone cannot establish target recognizability, constituent appearance, or overall perceptual quality.
Questions for authors
  1. Could you compare PartField with a geometry-based partition that preserves connectivity, using similar region counts and the same retrieval and selection pipeline?
  2. Could you provide rendered mosaics and side-by-side baseline comparisons, including close-ups and failure cases?
Reviewer_DvTYOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents PARTMosaic3D, a framework for 3D object mosaic synthesis from a fixed gallery that makes two decisions explicit: how the target is decomposed into regions and which gallery object is assigned to each region. Regions come from a PartField semantic-part hierarchy refined by a retrievability criterion (with optional HoloPart completion), while constituent selection minimizes an energy combining shape/appearance fidelity, placement fit, reuse, and pairwise interpenetration; the fidelity-and-reuse core is solved exactly by min-cost flow, and the full energy is solved with CP-SAT to certified optima within candidate pools. In controlled experiments on four textured targets with a 16,466-object gallery, parts outperform matched-granularity voxels, and joint selection improves results mainly through its geometric fit and overlap terms rather than through exact optimization of retrieval/reuse; retrievability-driven refinement and HoloPart completion show no benefit.

Critical strengths
  1. It makes the two decisions explicit and compares greedy and joint selection on the same energy. The paper separates “how to decompose the target into regions” from “which gallery object to assign to each region,” and formulates selection as a single energy. Proposition 1 shows that the fidelity-and-reuse core is solvable exactly by min-cost flow, and Proposition 2 shows that the baseline’s greedy reuse penalty is an order-dependent approximation of the same core energy. The full energy is solved with CP-SAT to certified optima within candidate pools. This is consequential because it turns the method from another heuristic pipeline into a fair, formal, and certifiable comparison framework; even if the empirical gains are modest, the framework itself is reusable.

  2. The controlled factorial experiments and honest negative results provide mechanistic insight. The paper controls region count, shares renderers and frozen features, fixes weights on Tiger, and transfers them to held-out targets. It separates the effects of parts vs. voxels and greedy vs. joint selection. It further finds that the benefit of parts comes mainly from where boundaries fall, not from retrievability; the benefit of joint selection comes mainly from the geometric fit and overlap terms, not from exactly optimizing fidelity/reuse; and retrievability-driven refinement and HoloPart completion show no benefit. This is consequential because the paper does not just report metric gains—it corrects plausible but wrong intuitions and gives clear design guidance for future 3D mosaic methods. The negative results also increase credibility.

Critical weaknesses
  1. There is essentially no qualitative visual evidence. Apart from Figure 1, which is only a pipeline overview, the paper provides no rendered final mosaics, no side-by-side comparisons with the voxel/greedy baselines, and no visual failure cases. For a 3D visual synthesis paper whose central claims are that “boundaries matter” and that “geometric terms improve the mosaic,” this is a serious omission. Coverage, F-score and Chamfer distance are proxies; without qualitative results, readers cannot judge whether the reported gains are visually meaningful, whether protrusion/interpenetration are actually removed, or how the method fails on fragmented targets such as Sports Car. This is consequential because it weakens the core evidentiary basis for acceptance: the paper asks readers to trust metric changes without showing what the resulting mosaics look like.

2.The empirical evidence is too narrow and the key metric is fragile. All conclusions rest on four targets, one run per condition, and weights chosen on Tiger; the held-out targets show consistency, not statistical significance or confidence intervals. Moreover, the interpenetration measure is only a lower bound, is sensitive to a single lattice pitch, and becomes unreliable near the lattice resolution—the paper itself reports a pitch reversal on Flamingos—while a large fraction of Sports Car constituents are too thin to be measured. Optimality is also certified only within small assignment pools (K_a=10). This is consequential because the paper’s central mechanistic claims—especially that joint selection helps through geometric terms and that overlap control works—may be driven by target-specific weight choices or metric artifacts rather than a robust, general effect.

Questions for authors
  1. Qualitative results. The paper appears to contain no rendered final mosaics or side-by-side visual comparisons with the voxel/greedy baselines; Figure 1 is only a pipeline overview. Can the authors provide qualitative figures for all four targets, including failure cases (especially Sports Car) and before/after geometric-term comparisons? If such figures exist, why were they omitted? This directly affects whether the reported metric gains are visually meaningful.

  2. Stability of weights and single-run evidence. Weights were selected on Tiger and then fixed for held-out targets, with one run per condition. Can the authors report whether the main conclusions (parts vs. voxels; joint vs. greedy; geometric terms vs. exact fidelity/reuse optimization) hold across a reasonable weight grid and across repeated runs with confidence intervals? This affects whether the findings are robust or target/weight-specific.

  3. Reliability of the interpenetration measure. The paper reports a pitch reversal on Flamingos and notes that many thin constituents are invisible to the measure. Can the authors show that the overlap-term conclusions hold under multiple pitches or a continuous/multi-resolution measure, or at least on the subset of constituents that are reliably measurable? This affects the central claim that geometric overlap control works.

  4. Meaning of “certified optimum” at K_a = 10. Optimality is certified only within small assignment pools, and Appendix D.8 reports that at K_a = 20 CP-SAT reaches a time limit with a gap of 0.08–0.11. Can the authors clarify how sensitive the conclusions are to K_a , and whether the “certified” results support claims about the full selection problem or only about the restricted pools? This affects how strongly the optimization contribution can be interpreted.

Suggestions for future work:

  1. Add a perceptual/user study and more targets with repeated runs.
  2. Evaluate a decomposition-free baseline and joint optimization of placements.
  3. Use multi-pitch or continuous interpenetration measures and verified real-world scale metadata.
  4. Test whether tuning the refinement threshold or HoloPart settings changes the negative results.
Reviewer_5qaiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

PartMosaic3D explores the approximation of a target 3D object by assembling complete objects retrieved from a repository. The approach decomposes the task into two decision-making steps: partitioning the target object into distinct regions and selecting appropriate constituent parts for each region. During the partitioning phase, the method leverages a hierarchical structure of semantic parts; for part selection, it employs joint optimization to balance fidelity, fit, reusability, and the costs associated with interpenetration between parts. However, the experimental evaluation provided is extremely limited (all experiments focus on only 4 objects), failing to demonstrate the method's generalizability. Furthermore, given that this is a computer vision task, the absence of any subjective evaluation raises doubts regarding the true validity and significance of the method's contributions.

Critical strengths

Clear formulation on methodology. The paper makes decomposition and constituent selection explicit, allowing their effects to be studied separately.

Critical weaknesses

The main proposed mechanisms receive mixed support. Retrieval-driven refinement does not outperform size-based refinement, and exact optimization of fidelity and reuse alone changes surface quality only marginally. The demonstrated gains mainly come from part boundaries and geometric costs.

The value of joint optimization is not fully isolated. The full joint method includes fit and overlap terms that are absent from the main greedy baseline. The comparison therefore does not establish how much improvement comes from joint optimization itself versus adding better geometric information.

Evaluation is limited and the overlap metric is unreliable for thin objects. Experiments cover only four targets, with one run per condition and no perceptual study. The voxel-based interpenetration measure misses thin constituents and shallow intersections; the paper also reports that its conclusions can depend on voxel resolution.

Limited evaluation and lack of qualitative and perceptual validation. The experiments use only four target objects, with Tiger also used for hyperparameter selection. Such a small test set is insufficient to establish generalization across object categories and geometric structures. No qualitative results are shown in the paper either.

Questions for authors
  1. Can a greedy or approximate method using the same full objective achieve comparable results?

  2. Why retain retrieval-driven refinement as a main contribution when size-based refinement performs similarly or better?

  3. How sensitive are results to the candidate-pool size, solver budget, and fixed placement strategy?

  4. Do the findings generalize to more targets and align with human judgments of mosaic quality?

  5. Can the authors provide multiview renderings of the generated mosaics alongside the target objects and baseline results?

  6. Can the authors conduct a user study on target recognizability, constituent identifiability, and visual quality?

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

Related papers

Discussion (0)

Sign in to comment.