Learning Fine-to-Coarse Cuboid Shape Abstraction

Gregor Kobsik; Isaak Lim; Leif Kobbelt; Morten Henkel; Tim Elsner; Victor Czech; Yanjiang He

REVIEW 3 major objections 5 minor 3 references

Learning Fine-to-Coarse Cuboid Shape Abstraction

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A fine-to-coarse training schedule with volume preservation yields 3D cuboid abstractions that are both more compact and more accurate than previous methods.

desk verdict Fine-to-coarse cuboid abstraction with a volume-preservation loss is a genuinely new and effective recipe, but the reported gains lean on an undocumented volumetric ground truth and on thin statistics. read the letter →

arxiv 2502.01855 v1 pith:7WC3GNUS submitted 2025-02-03 cs.CV cs.GR

Gregor Kobsik , Morten Henkel , Yanjiang He , Victor Czech , Tim Elsner , Isaak Lim , Leif Kobbelt This is my paper · ORCID

classification cs.CVcs.GR

keywords shapeabstraction3Dstructuralrepresentationco-segmentationcuboidprimitivesfine-to-coarsetrainingvolumepreservationpointcloudunsupervisedlearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that cuboid abstraction of 3D shapes improves when the network starts with a deliberately overcomplete set of primitives and gradually prunes them during training, rather than fitting a fixed small number from the outset. The authors add a volume-preservation loss on top of the usual surface Chamfer loss, which prevents thin degenerate cuboids and improves intersection-over-union with the original volume. On ShapeNet and DFAUST categories, they report the best or second-best Chamfer distance and IoU while using the fewest cuboids among cuboid-based baselines. If correct, unsupervised structural abstraction becomes accurate enough to support co-segmentation, retrieval, and symmetry detection without manual annotations.

What carries the argument

The load-bearing mechanism is the fine-to-coarse schedule, implemented by an abstraction loss that forces the number of active cuboids toward a prescribed target via binary cross-entropy on existence probabilities, plus a reconstruction loss that combines surface Chamfer distance with a volume-preservation Chamfer term sampled from cuboid interiors. The architecture processes the input point cloud through two Vision Transformers: the first builds global shape features from local point patches, and the second lets learnable cuboid latents attend to those features to predict each cuboid's rotation, translation, scale, and existence probability. During inference, a merging post-process replaces highly overlapping cuboids with the oriented bounding box of their union, further reducing the primitive count without substantial loss of accuracy.

What would settle it

Train the same architecture and schedule on a dataset that offers only surface point clouds (no interior samples) and evaluate IoU on the test set; the paper's own ablation predicts that the cuboids degenerate to thin plates and IoU collapses toward zero. If the method still maintains high IoU under purely surface supervision, the central claim about volume preservation would be overturned.

Watch

Extended reading notes

Core claim

The central claim is that a fine-to-coarse training scheme, combined with a reconstruction loss that enforces volume preservation as well as surface approximation, produces cuboid shape abstractions that are simultaneously more compact and more faithful than those of prior work. The network begins with 128 cuboids and is driven to a user-specified minimum count through a binary cross-entropy abstraction loss applied to each cuboid's existence probability; this lets it lock in fine details early and then discard or merge redundant primitives. A volume loss, computed against interior point samples, prevents the degenerate thin cuboids that surface-only losses allow. In the plane, chair, table, and human categories, the method achieves the best or near-best Chamfer distance and IoU while using the least or second-least number of cuboids, and the resulting cuboids support co-segmentation, clustering, retrieval, and partial symmetry detection.

Load-bearing premise

The method assumes that volumetric point clouds of the training shapes are available to compute the volume-preservation loss; the paper never specifies how these interior samples are generated, and removing the volume loss drops IoU to zero in the ablation.

Editorial extensions

If this is right

Unsupervised cuboid abstraction can reach state-of-the-art surface and volume fidelity while using fewer primitives, making the representation more useful for structure-aware modeling and editing.
The fine-to-coarse schedule is a transferable strategy for primitive-fitting networks and could extend to superquadrics or deformable templates by changing the primitive parameterization.
Because inference needs only a surface point cloud, the trained model can abstract new shapes on demand without volumetric supervision.
The learned cuboids carry enough structural signal to support co-segmentation, clustering, retrieval, and partial symmetry detection without annotation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The method's reported gains depend on volumetric point clouds for the volume loss, which are not available from surface-only scans; a practical deployment would need a way to estimate or predict interior points.
The merge post-processing could be reformulated as a differentiable loss or a learned selection module, allowing the network to predict the final compact abstraction in a single forward pass.
The fine-to-coarse schedule resembles a curriculum on primitive count; the same principle might improve other structured prediction tasks such as skeleton extraction or part-graph discovery.
Since the paper notes that re-indexing merged cuboids could improve co-segmentation, the current indexing scheme likely underestimates the method's semantic consistency.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a solid, useful paper that deserves refereeing, and the main caveat is exactly where the stress-test points — the volume point cloud X_V is never specified.

What's new and good: unlike HCA (coarse-to-fine), CAS, and Tulsiani (fixed counts), the network starts with 128 cuboids and smoothly prunes to ~7–10 during training. The abstraction loss with binary cross-entropy on existence probabilities is a clean way to steer the count. The volume-preservation loss L_vol is new in this cuboid setting, and the ablation (Table 3) makes a convincing case that it is the difference between IoU 61 and IoU 0, and between sensible cuboids and degenerate slivers. The merging post-process is straightforward, and the threshold analysis (Table 4) honestly shows the trade-off. Reported CD/IoU numbers are directionally better than CAS/HCA at comparable or lower primitive counts on most classes. This is a genuine step for compact structural abstraction.

Soft spots, in proportion. The missing X_V generation is the biggest. The method needs volumetric point clouds for training, but Section 4 never says how they are obtained — mesh voxelization, interior sampling, watertightness? Table 3 shows removing L_vol collapses IoU to 0.0, so the headline advantage rides on it. The Limitations section does not mention this dependency. That is a reproducibility hole, not a fatal flaw, but it needs to be fixed by specifying the procedure or releasing code and data.

Statistics are thin. Baselines are selected as best-of-five runs; the authors do not report variance for their own runs, and the merge threshold is tuned per class. The abstract says “fewer cuboids than previous work,” but Table 1 shows they use fewer than CAS and DPF while using more than HCA on plane and chair. That overstatement should be toned down. Also, the “unsupervised” label is fair only in the sense that no part labels are used; training requires volumetric occupancy ground truth, which is a form of dense annotation. A sentence of clarification would help.

Minor: the downstream sections (clustering, retrieval, symmetry) are qualitative demonstrations — fine for a graphics paper but not evidence.

Who this is for: anyone working in shape abstraction, structural co-segmentation, or primitive-based reconstruction. The method is practical, the writing is clear, and the ablations are informative. It deserves peer review; a careful referee should ask for the X_V procedure, error bars, and code. I would bring it to reading group in current form, mostly to discuss what “unsupervised” really means here.

Recommendation: engage, with revision expectations.

Referee Report

3 major / 5 minor

Summary. The paper proposes a fine-to-coarse unsupervised learning method for cuboid shape abstraction. The network is trained to predict a large set of cuboids and then gradually reduces their number through a loss that penalizes redundant primitives, subject to a user-specified target count. A second contribution is a volume-preserving reconstruction loss that complements the surface Chamfer loss. The method is evaluated on ShapeNet classes (plane, chair, table) and a DFAUST subset, where it reports improved Chamfer distance and IoU relative to HCA, CAS, and DPF-PPM, often with fewer cuboids. The abstraction is also applied to co-segmentation, clustering, retrieval, and partial symmetry detection.

Significance. If the results hold, the fine-to-coarse training schedule is a worthwhile alternative to fixed-primitive-count optimization, and the explicit volume-preservation loss directly addresses a known degeneration of cuboid abstractions (thin surface shells). The paper provides user-controlled compactness and demonstrates downstream utility. It is, however, an empirical paper without code or released models, and its central claims rest on an undocumented volumetric supervision signal and on a benchmark protocol that does not fully report variance. These issues are fixable and do not invalidate the approach, but they must be addressed before the state-of-the-art claims can be assessed.

major comments (3)

[Section 3.2.1 (Eq. 6) and Section 4 (Datasets)] The volume point cloud X_V is never defined. The paper states in Section 3 that the network is trained with 'a surface point cloud X_S and a volume point cloud X_V', but Section 4 does not specify how X_V is generated for ShapeNet or DFAUST, how many interior points are sampled, whether meshes are made watertight beforehand, or whether this information is available at inference time. This is load-bearing: Table 3 shows that removing L_vol (L_rec = L_surf) collapses IoU to 0.0 and increases CD from 0.024 to 0.058 on the plane class, and the paper attributes the avoidance of degenerate thin cuboids to L_vol. The claim that the method is 'unsupervised' is also overstated, because training requires volumetric occupancy ground truth (interior samples) even though part labels are not used. Section 4.6 does not disclose this dependency. Please provide the exact X_V generation procedure, state its availability requirements, and either re-word 'unsupervised' or justify it as unsupervised with respect to semantic labels.
[Section 4 (Reference methods) and Section 4.1 (Table 1)] The comparison protocol is asymmetric. For the baselines, the paper states (Section 4, Reference methods): 'we train the method five times for every category using the standard hyper-parameters and report the best performing run in terms of Chamfer distance.' For the proposed method, no such multi-seed protocol is reported for the main results in Table 1, and no variance or significance tests are given. On several key entries the reported advantage is small or zero (e.g., chair CD: Ours 0.036 vs CAS 0.036; chair Num: Ours 8.37 vs CAS 9.77, but the CD tie means the 'more precise' claim is not supported there). The paper should report mean and standard deviation over at least five runs for all methods, or clearly state that the same best-of-five selection was applied to both the baselines and the proposed method.
[Section 3.3 and Table 4] The merge threshold theta_merge is a per-class tuned post-processing parameter (plane 1.2, chair 1.4, table 1.0, human 1.0). Table 4 shows that on the plane class, increasing theta_merge monotonically improves both CD and IoU, and that the 'w/o' row (no merging) gives the best CD (0.024) and IoU (61.1) of all configurations. Thus the reported 'best scores ... while using least cuboids' configuration is obtained by deliberately sacrificing reconstruction quality to reduce primitive count, and the choice of theta_merge is not derived from a principled criterion. The paper should either present the full trade-off curve and recommend a criterion, or show that the chosen theta_merge is stable across classes and random seeds rather than selected per benchmark.

minor comments (5)

[Section 4.1, Table 1 caption] The caption says 'Intersection of Union' but should be 'Intersection over Union'.
[Section 4.5] There is a typo: 'learnarble cuboid latents' should be 'learnable cuboid latents'.
[Figures 5 and 6 captions] The captions use 'distinc parts' instead of 'distinct parts'.
[Section 3.2.1, Eq. (3)] The notation gamma_bar_m is confusing: the sentence 'gamma_bar_m is a shorthand notation to denote that there exists a primitive closer than primitive m' does not align with the product over m-bar < m in Eq. (3). Please clarify the ordering and the exact meaning of gamma_bar_m.
[Section 4.5, Table 3] The row 'w/o l_c' reports Num=37.32 and moderate CD/IoU, but the text says 'training without using l_c explicitly diverges.' If the configuration diverges, it would be clearer to state that the reported metrics correspond to the best of five runs that did not diverge, or to omit the row and simply state the failure.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: held-out test evaluation, user-specified primitive count, and no load-bearing self-citation.

full rationale

This is an empirical learning paper whose claims are supported by evaluation on held-out test shapes, not by a derivation that reduces to its own inputs. The fine-to-coarse scheme uses a user-specified target function Γ(φ) (Eq. 8) and an abstraction loss L_abs (Eq. 9) to control the number of primitives; because the paper explicitly presents this as user control rather than as an emergent prediction, the reported primitive counts are a designed property, not a circularly derived result. The reconstruction loss combines a surface Chamfer term (Eq. 5) with a volume Chamfer term (Eq. 6) trained on volumetric point clouds X_V, while the IoU benchmark is computed separately by grid sampling at 128^3 on held-out shapes, so the evaluation metric is not identical to the training loss and is not forced by construction. The paper does omit the procedure for generating X_V, which is a reproducibility and supervision-disclosure weakness and makes the 'unsupervised' label potentially misleading, but it is not a circularity: X_V is another representation of the same input shape, not a fitted parameter renamed as a prediction. The only self-citation (Lin et al. 2020, cited in Section 2.1 as related primitive-fitting work) is not load-bearing, and no uniqueness theorem or ansatz is imported from the authors' prior work. Ablations such as L_rec = L_surf dropping IoU to 0.0 (Table 3) show expected dependency on the volume loss, but this is an empirical ablation, not a circular step.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claim depends on a set of hand-chosen hyperparameters (loss weights, target primitive counts, merge thresholds) and on the availability of volumetric ground truth during training. These are reasonable in the benchmark setting but limit the generality of the method. No new physical or mathematical entities are introduced.

free parameters (9)

kappa_max (initial cuboid count) = 128
Maximum number of cuboids at the start of training; hand-chosen. Section 4 hyperparameters.
kappa_min (target final cuboid count) = 7 (plane/table/human), 10 (chair)
User-specified target number of primitives at the end of training; the central 'compact abstraction' claim depends on this choice. Section 4 hyperparameters.
lambda_abs = 1e-3
Weight of abstraction loss controlling the trade-off between reconstruction and compactness; hand-tuned. Section 4.
lambda_vol = 1e1
Weight of volume preservation loss; hand-chosen. Section 4.
lambda_surf = 1e0
Weight of surface reconstruction loss; hand-chosen. Section 4.
theta_merge = 1.2 (plane), 1.4 (chair), 1.0 (table), 1.0 (human)
Merging threshold for redundant cuboids, tuned per class; Table 4 shows it strongly affects the reported Num and IoU. Section 4.
epsilon_size = not stated
Small inflation factor used when estimating cuboid volume ratios for merging; value not reported. Section 3.3.
alpha, beta (Chamfer weights) = 1.2, 0.8
Weights for P->X and X->P terms in the Chamfer loss, taken from Paschalidou et al. 2019; not tuned here. Section 3.2.1.
N (local tokens) and K (neighbors) = 128, 32
Number of local tokens and nearest neighbors in the encoder; design choices. Section 3.1.1.

assumptions (5)

domain assumption The input shapes are axis-aligned; the network is not rotation-invariant.
The limitation section states 'we assume our dataset to be axis-aligned.' This is load-bearing because a rotated shape would likely receive a different abstraction. Section 4.6.
domain assumption Volumetric point clouds X_V of the training shapes are available to compute L_vol.
The volume loss is defined in Eq. (6) against X_V, but the paper does not specify how X_V is generated (e.g., mesh volume sampling). The IoU advantage in Table 1 would not transfer to settings without volumetric supervision. Section 3.2.1.
ad hoc to paper The BCE abstraction loss on existence probabilities gamma reliably enforces the target primitive count and preserves the most useful primitives.
The loss is defined in Eq. (9) and is the mechanism that reduces the primitive count, but the paper provides no analysis showing that the top-gamma primitives are the correct ones to keep; the ablations only vary the loss weight. Section 3.2.2.
ad hoc to paper The volume-ratio criterion V_ratio identifies overlapping cuboids that can be merged without significant reconstruction loss.
The merging post-processing is introduced in Section 3.3 with a heuristic threshold; the choice of threshold per class affects the final primitive count and quality. No theoretical guarantee is given.
domain assumption Chamfer distance between sampled point sets is a suitable surrogate for reconstruction quality.
The reconstruction loss and evaluation metrics use Chamfer distance; this is a standard choice in the field, but it is an assumption that the metric correlates with perceived or structural quality. Section 3.2.1.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning Fine-to-Coarse Cuboid Shape Abstraction." pith.science (2026). https://pith.science/paper/7WC3GNUS

@misc{pith2026250201855,
  author       = {Pith},
  title        = {Pith review of: Learning Fine-to-Coarse Cuboid Shape Abstraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7WC3GNUS}},
  note         = {Machine review of arXiv:2502.01855}
}

read the original abstract

The abstraction of 3D objects with simple geometric primitives like cuboids allows to infer structural information from complex geometry. It is important for 3D shape understanding, structural analysis and geometric modeling. We introduce a novel fine-to-coarse unsupervised learning approach to abstract collections of 3D shapes. Our architectural design allows us to reduce the number of primitives from hundreds (fine reconstruction) to only a few (coarse abstraction) during training. This allows our network to optimize the reconstruction error and adhere to a user-specified number of primitives per shape while simultaneously learning a consistent structure across the whole collection of data. We achieve this through our abstraction loss formulation which increasingly penalizes redundant primitives. Furthermore, we introduce a reconstruction loss formulation to account not only for surface approximation but also volume preservation. Combining both contributions allows us to represent 3D shapes more precisely with fewer cuboid primitives than previous work. We evaluate our method on collections of man-made and humanoid shapes comparing with previous state-of-the-art learning methods on commonly used benchmarks. Our results confirm an improvement over previous cuboid-based shape abstraction techniques. Furthermore, we demonstrate our cuboid abstraction in downstream tasks like clustering, retrieval, and partial symmetry detection.

Figures

Figures reproduced from arXiv: 2502.01855 by the authors.

**Figure 1.** Left: During training the number of primitives decreases from hundreds of cuboids to only a few ones. Colors are based on a global cuboid index. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Architecture overview: a surface point cloud [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Partial symmetry detection: Using our cuboid abstraction to decom [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: Left: k-means clustering (𝑘 = 5) on top of our cuboid parameters [𝑟, 𝑡, 𝑠,𝛾 ]. Each column represents a cluster. Right: Retrieval of the nearest neighbors from a collection of shapes. Top row shows the query cuboids. our model is able to recover geometric parts aligned…

**Figure 5.** Figure 5: Qualitative samples of the plane and chair class. We visualize the ground truth (GT) shape and the cuboid abstraction of [Sun et al [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 6.** Figure 6: Qualitative samples of the table and human class. We visualize the ground truth (GT) shape and the cuboid abstraction of [Sun et al [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 1 canonical work pages

[2]

ACM Transactions on Graphics (TOG) 38, 6 (2019)

SDM-NET: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG) 38, 6 (2019). Jiaxin Li, Hongxing Wang, Jiawei Tan, and Junsong Yuan. 2024b. Shared Latent Mem- bership Enables Joint Shape Abstraction and Segmentation with Deformable Su- perquadrics. IEEE Transactions on Image Processing (2024). Jun Li, Kai Xu, Siddharth...

work page 2019
[2017]

ACM Transac- tions on Graphics (TOG) 36, 4 (2017)

Grass: Generative recursive autoencoders for shape structures. ACM Transac- tions on Graphics (TOG) 36, 4 (2017). Songlin Li, Despoina Paschalidou, and Leonidas Guibas. 2024a. PASTA: Controllable Part-Aware Shape Generation with Autoregressive Transformers. arXiv preprint arXiv:2407.13677 (2024). Cheng Lin, Lingjie Liu, Changjian Li, Leif Kobbelt, Bin Wan...

arXiv 2017
[2019]

In Proceedings of the IEEE/CVF International Conference on Computer Vision

Bae-net: Branched autoencoder for shape co-segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision . Laurent Chevalier, Fabrice Jaillet, and Atilla Baskurt. 2003. Segmentation and su- perquadric modeling of 3D objects. (2003). Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at sc...

arXiv 2003

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [2]

ACM Transactions on Graphics (TOG) 38, 6 (2019)

SDM-NET: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG) 38, 6 (2019). Jiaxin Li, Hongxing Wang, Jiawei Tan, and Junsong Yuan. 2024b. Shared Latent Mem- bership Enables Joint Shape Abstraction and Segmentation with Deformable Su- perquadrics. IEEE Transactions on Image Processing (2024). Jun Li, Kai Xu, Siddharth...

work page 2019

[2] [2017]

ACM Transac- tions on Graphics (TOG) 36, 4 (2017)

Grass: Generative recursive autoencoders for shape structures. ACM Transac- tions on Graphics (TOG) 36, 4 (2017). Songlin Li, Despoina Paschalidou, and Leonidas Guibas. 2024a. PASTA: Controllable Part-Aware Shape Generation with Autoregressive Transformers. arXiv preprint arXiv:2407.13677 (2024). Cheng Lin, Lingjie Liu, Changjian Li, Leif Kobbelt, Bin Wan...

arXiv 2017

[3] [2019]

In Proceedings of the IEEE/CVF International Conference on Computer Vision

Bae-net: Branched autoencoder for shape co-segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision . Laurent Chevalier, Fabrice Jaillet, and Atilla Baskurt. 2003. Segmentation and su- perquadric modeling of 3D objects. (2003). Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at sc...

arXiv 2003