acceptodds
Under review as a conference paper at ICLR 2027

PixelDec: Pixel-aligned Superquadric Decomposition from a Single Image

Abstract

We present PixelDec, an approach to obtain pixel-aligned superquadric decompositions from a single RGB image. Previous approaches for superquadric decomposition have shown how a compact representation of 3D scenes can be useful for different downstream tasks, such as robotic navigation and path planning. However, all existing approaches are deterministic and tackle the problem from a reconstruction perspective, representing with primitives only the observed parts of the objects, and often yielding incomplete and fragmented decompositions. In this work, for the first time, we tackle primitive-based decomposition from a generative perspective, using the compact representation not only as our final output, but also as a compact representation space in which we efficiently learn the distribution of real-world 3D shapes. We train our model on ShapeNet and finetune it on Aria Synthetic Environments, and show that when conditioned on single-view inputs, PixelDec strongly outperforms previous decomposition methods as well as multi-stage pipelines built on state-of-the-art single-image reconstructors. Finally, we show how our method can be practically used for a range of downstream applications.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Vision, Superquadrics, Flow Matching, Generative Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pdB3OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PixelDec, a conditional flow-matching framework for generating superquadric decompositions from single-view observations. By distilling knowledge from a SuperFlex teacher that operates on complete point clouds, PixelDec aims to produce compact geometric representations while inferring unobserved object geometry. The authors evaluate the proposed method on ShapeNet, ASE, and TableVerse, examining both reconstruction quality and computational efficiency. The paper further explores its potential applications in downstream robotic tasks, including grasping, motion planning, and VLM-guided manipulation.

Critical strengths
  1. The paper presents a comprehensive generative pipeline for superquadric scene decomposition from single-view observations, aiming to produce compact geometric representations while inferring unobserved geometry.
  2. The method thoughtfully addresses the heterogeneous nature of superquadric parameters by introducing tailored flow-matching formulations for both continuous geometric parameters and discrete existence variables.
  3. The downstream table-decluttering experiments show encouraging improvements in grasp success rates, while the VLM-guided application provides an interesting illustration of how the learned representation can facilitate planning for physical interactions. I also appreciate the supplementary real-world demonstrations, which further highlight the practical potential of the proposed approach.
Critical weaknesses
  1. I would appreciate a clearer discussion of how the proposed algorithmic contributions differ from and extend prior work:
    • Prior work [1] has already explored generative modeling in the superquadric parameter space. While this paper extends this direction to reconstruction from partial observations, I wonder if the authors could further clarify the methodological contributions beyond this task extension, as the additional algorithmic novelty seems somewhat limited from my understanding.
    • PixelDec is trained under the supervision of a frozen SuperFlex teacher that operates on complete point clouds. While the improvements over partial-input baselines provide encouraging evidence of its completion capability, I wonder whether these results necessarily imply an improvement in decomposition quality under complete observations. From my understanding, the main contribution appears to be more closely related to conditional completion rather than advancing the underlying decomposition capability itself. It would be helpful if the authors could further clarify this distinction.
    • I appreciate the point-cloud-only comparison on ShapeNet, which helps ensure input fairness. However, the robotics evaluation additionally involves RGB conditioning and selection among multiple generated samples. I wonder if the authors could clarify how much of the improvement stems from generative modeling itself rather than these additional factors. Taken together, the results demonstrate useful practical benefits, while leaving some uncertainty about the extent of the independent algorithmic advance.
  2. Further analysis of the sources of improvement would be helpful. The ablation studies provide useful insights into image conditioning, existence modeling, and sampling steps. However, I wonder if the authors could further clarify a few remaining questions, such as:
    • A controlled comparison with deterministic regression under similar architectural settings, conditioning, and teacher supervision would help better understand the specific benefits of generative modeling.
    • A controlled evaluation under varying degrees of input occlusion or truncation would help better understand the scenarios in which generative modeling is most beneficial.
  3. The evaluation of computational efficiency could benefit from further clarification and more comprehensive reporting.
    • The meanings of Memm\text{Mem}_m and Mems\text{Mem}_s in Tab.2 are unclear to me. It would be helpful if the authors could specify the hardware used for runtime evaluation to improve the interpretability and reproducibility of the results.
    • The single-image pipeline relies on instance segmentation and point-map prediction using SAM3 and MoGe2. From my understanding, the runtime reported in Tab. 2 may not include these preprocessing costs, while the robotics timing starts from already segmented point clouds. I wonder if the authors could clarify the end-to-end latency, including preprocessing, to better support the real-time claim in Sec. 4.3.

[1] Zhiyang Liu et al. Rethinking 3D Shape Generation: Diffusion over Superquadrics. In ICML, 2026.

Questions for authors
  1. It would be helpful to clarify the methodological novelty and isolate the benefits of generative modeling.
  2. Ablations comparing deterministic regression with generative modeling, alongside evaluations across different levels of occlusion or truncation, could help clarify the benefits of generative modeling.
  3. Clarifying the memory metrics, evaluation hardware, and end-to-end computational costs would strengthen the efficiency analysis.
  4. I wonder if Fig. 7 could be presented more clearly, as I find it somewhat difficult to identify the attention patterns between semantically similar or structurally dependent parts discussed in the paper.
  5. It would be helpful if the authors could include some generated 3D assets (e.g., .glb files) from both the proposed method and the baselines in the supplementary material to facilitate qualitative comparison.
6 comments

We thank the reviewer for their thoughtful and honest feedback. Below, we address the individual questions with more in-depth explanations. We are happy to provide further explanations and additional results if needed.

1. Comparison to Prior Work The reviewer asks how our approach differs from “Rethinking 3D Shape Generation: Diffusion over Superquadrics” (DoSs) [ICML’26]. We thank the reviewer for pointing us to this work, which we will discuss in the revised paper. Both methods share the high-level idea of generative modeling directly in superquadric parameter space. However, they follow different algorithmic approaches, and PixelDec operates in a substantially more challenging setting, enabled by several methodological contributions. We discuss these differences below.

  1. Class-agnostic vs. class-specific. PixelDec is class-agnostic, while DoSs is not: DoSs primarily reports results of separate models trained for each object category, and is demonstrated on only three categories (chair, airplane, and car). PixelDec requires no class labels: a single model is trained jointly on all available objects. This is essential for our setting, since PixelDec must reconstruct arbitrary objects in scenes whose object categories are not known in advance, and it makes the problem considerably harder, as one model must handle a far more diverse distribution of shapes. In comparison, a category-specific model such as DoSs only needs to capture the shape variation within one category, which is comparatively easy: airplanes, for example, mostly consist of a fuselage, a pair of wings, and a tail.
  2. Object pose. DoSs generates all shapes in a canonical object frame: while each superquadric has its own pose, shapes are always generated with the same global orientation, e.g., the seat of a chair is always horizontal and its backrest always faces the same direction. This considerably simplifies the generation problem. Instead, PixelDec generates all primitives directly in the observation camera coordinate frame, in which objects appear at arbitrary orientations relative to the camera. The model must therefore infer each object’s orientation from the input image and point map, in addition to its shape.
  3. Unconditional vs conditional generation. DoSs performs unconditional or category-conditional generation: beyond selecting one of three categories (car, airplane, or chair, through model selection or a one-hot label), the generated shape is not tied to any observation. Moreover, the generated objects are always placed in the same canonical pose, with the same scale and orientation and centered at the origin (see previous point). PixelDec is considerably more general: it generates arbitrary shapes, namely the specific object depicted in the conditioning image, without being restricted to predefined categories, and places it correctly in the camera coordinate frame, consistent with the input observation.
  4. Joint continuous-discrete flow vs. continuous diffusion. DoSs models all superquadric parameters as continuous variables. PixelDec instead combines different types of flows: continuous (euclidean, and riemann manifolds for rotations - see Tab. 7), and discrete flow (for primitive existence).
  5. Deformed superquadrics vs regular superquadrics. DoSs builds on Marching-Primitives [CVPR’23] and uses regular superquadrics, while PixelDec builds on SuperFlex [ECCV’26] and uses deformable superquadrics with tapering and bending. This enables improved compression: PixelDec represents objects with only 16 primitives compared to 128 for DoSs.

Taken together, PixelDec differs from DoSs both in its setting (class-agnostic, conditional, pixel-aligned generation) and in its method (deformable superquadrics and joint continuous–discrete flow). As a result, PixelDec reconstructs complex real-world scenes, runs in working robotic systems, and matches or outperforms state-of-the-art 3D generative methods at a fraction of their runtime and memory cost, whereas DoSs only demonstrates generation of three object categories in canonical pose.

2. Benefit from Generative Modeling, RGB/DINO features, and sample selection
The reviewer asks for a comparison of deterministic regression with generative modeling. We trained a deterministic version of PixelDec (Deterministic student) and we report qualitative results in Fig.18 (revised paper) and quantitative in the table below. These results overall show that the impact of using a generative approach is more significant than mere RGB features. We thank the reviewer for the inspiration for these experiments and we will include them in the revised version of the paper.

MethodRGBGen.IoU ↑L1 ↓L2 ↓P-FID ↓P-KID ↓# Prim.
SuperFlex✗✗0.4632.340.1512.228.64.91
Deterministic Student✓✗0.4472.590.153.926.404.35
PixelDec✗✓0.6031.700.080.880.135.58
PixelDec✓✓0.6221.620.060.860.065.59

In robotic experiments we select the best sample among multiple generations. In practice, we select the sample with lowest point-to-surface distance (P→SQ, computed on observed points). The P→SQ metric unsurpirsingly improves when compared to the average one (see results below), while other metrics do not change substantially. We think the sample selection is not crucial to our method.

PixelDec samplingIoU ↑Chamfer-L1 ×100 ↓P→SQ ↓
Average over 10 draws0.5005.802.55
Draw with lowest P→SQ0.5105.521.89

3. Evaluations across different levels of occlusion or truncation.
The reviewer is curious of how our method performs under different levels of occlusion. We report qualitative results in Fig.19 and quantitative in Fig. 20 of the updated paper. They show that our PixelDec generates good decomposition also for strongly occluded objects while this is not the case for baselines. Notice that in the figure we also include the new Deterministic Student related to the point above.

4. Memory, hardware, and computational costs.
The reviewer asks for details on our evaluation setup around memory usage, hardware and computation cost of all method components.

  • In Table 2, Mem(_m) denotes GPU memory occupied by model weights, while Mem(_s) denotes additional GPU memory required for single-object inference.
  • Runtimes in the main paper are measured on an NVIDIA A100 (80 GB) with models already loaded, excluding SAM3 instance segmentation and MoGe2 point-map prediction, which is shared among all the methods. Preprocessing takes 0.430 s (SAM3: 0.367 s with batched prompts; MoGe2: 0.063 s). We include runtimes for SAM3, MoGe2 in Fig. 17 (left) in the updated paper.
  • Fig. 17 (middle and right) also provides clearer runtimes comparison of different methods on Big Bang Theory scene (middle) and in function of number of objects (right), computed on an RTX 4090. SAM3D and ShapeR runtime grows linearly with the number of objects, as their inference already saturates GPU compute. In contrast, PixelDec reconstructs all objects in a single batch, maintaining nearly constant runtime as the number of objects increases.

5. Does PixelDec improve decomposition quality of SuperFlex under complete observations?
SuperFlex is the teacher which defines the upper bound of the decomposition quality of PixelDec. PixelDec itself does claim contributions with regard to decomposition quality.

Q5. Provide GLB files. Thank you for the suggestion. We added an interactive HTML viewer to the supplementary material to compare 3D reconstructions from PixelDec and the baselines side by side. After extracting the supplementary ZIP, open interactive_html_viewer/index.html in a browser. Standalone .glb files for PixelDec and the baselines are included in interactive_html_viewer/assets/. Please let us know if you have any trouble opening the viewer or accessing the visualizations.

Reviewer_D2bcOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes generating pixel-aligned superquadrics from a single RGB image to represent the objects in a scene. With this method, the paper can recover the complete superquadrics of masked objects from the incomplete observation (the image and its partial point map). This recovered result can benefit practical applications such as robotic grasping.

Critical strengths
  1. The superquadric representation generated by this paper is very lightweight (5.9 kB for a whole scene), and its generation speed is very fast (0.07 s per object).
  2. The paper can recover the complete object shape from the image and its partial point cloud, and since the predictions are made in the frame of the input point map, the results are directly placed back into the original 3D scene without a separate pose estimation step.
Critical weaknesses
  1. Do the ASE and TableVerse datasets separate the training set and the test set? The paper does not explain how they are split. In addition, some cases in Fig. 4 are very difficult. For example, in the first two columns, the input point clouds only cover a small part of the objects, but PixelDec still generates complete and detailed objects. How is this performance obtained? The authors need to provide an explanation.
  2. The image is cropped around the predicted instance mask and fed into DINOv2. For occluded or truncated objects, is only the image of the visible part used as input, or is a prior, such as the complete bounding box size of the object, used as the image size to remind the network that this is a complete object? More importantly, the 3D points of each instance are mean-centered and scaled into the unit cube based only on the observed part. If only half of an object is observed, after this normalization, how can the network recover the complete object, including its correct extent and scale? The paper does not explain these points clearly, and I think they are important for the complete recovery, because this problem does not exist on the ShapeNet dataset.
  3. The paper shows a clear performance drop on the real indoor scene results. For example in Fig5, the bookshelf in the back of the first row, and the drawer on the right that is occluded by the pillar (green prediction). It does not seem to solve the occlusion problem, and the authors need to provide an explanation.
  4. In the real-robot experiments, including the demo videos, the reconstruction results look poor. The mug’s handle is not predicted, and the two handles on the sides of the pot are not predicted either. Could the authors explain these results? This makes me question whether the method can really be used in practical scenarios. The paper does not show good results in either visual quality or geometric fidelity.
Questions for authors
  1. Explain the experiment setting on ASE and TableVerse datasets.
  2. Explain how the images and pc cropped and normalized.
  3. Explain the performance on real and robot experiments.
  4. As the poor results of cup and pot in robot experiments, they can not be used for robot interaction experiments. It would be helpful if author could further explain the potential application of PixelDec besides effeciency.
8 comments

We thank the reviewer for their thoughtful feedback. It appears the reviewer has significant concerns about how well our method works in practice. We hope that our answers below address these, and we are happy to provide further explanations and additional results if needed.

1.1 Experimental Details on ASE and TableVerse
The reviewer is wondering whether we have a separate train and test set on ASE and TableVerse. We confirm that ASE and TableVerse have separate training and test splits, and no test sample was ever seen during training or fine-tuning.

  • For ASE, whose scenes are populated with objects from the ABO dataset, we select test scenes containing only ABO objects that never appear during training of our method, ensuring separation at both the scene- and object-levels. We evaluate on 300 distinct held-out objects across 220 test scenes.
  • For TableVerse, whose object meshes are generated individually using SAM3D, we split by scene ID and evaluate on 276 held-out scenes. We will include these details in the revised paper.

In addition we evaluate all methods zero-shot on two real-world datasets that none of the methods were trained or fine-tuned on and report it in the revised paper. We hope that these explanations and additional experiments provide sufficient quantitative evidence for our method performance in general real-world scenarios.

1.2 Performance of Occluded Objects
The reviewer asks how PixelDec is able to reconstruct strongly occluded objects more accurately than the baselines.

First, we want to emphasise that our approach does not have access to privileged information like the complete object bounding boxes as asked by the reviewer (also see next paragraph on ‘cropping and normalization’).

Second, we explain the qualitative results in Figure 4 which illustrates how our method can reconstruct a lamp (1st column) and a chair (2nd column) that are strongly occluded, while the ShapeR and SAM3D struggle to do so. We attribute the limited performance of SAM3D and ShapeR on those examples to architectural design choices of the methods:

  • ShapeR can only reconstruct geometry extending ~10% beyond the bounding box of the observed points. This is because each object is normalized so that its bounding box fills a fixed cube, and both the generative model and the VAE decoder are trained only on shapes within that cube. When occlusions leave large parts of an object unobserved, ShapeR therefore cannot recover them, even if it correctly infers the object semantics.

  • SAM 3D generates the object’s shape in a canonical frame and predicts its pose (rotation, translation, scale) as a separate output. We observe that, under strong occlusion, this can produce a plausible shape but an incorrect pose. For the lamp, the shade is reconstructed well but wrongly rotated by 90°.

  • PixelDec avoids both limitations. Since it generates superquadric parameters directly, without a fixed decoding volume, its output is not bounded by the extent of the observed points. It also does not separate a canonical shape from an object-level pose, as SAM 3D does. Instead, each primitive’s position, orientation and size are generated directly in the observation frame, so shape and placement are predicted together.

We will include this discussion in the revised paper.

2. Cropping and Normalization
The reviewer asks how inputs are normalized (specifically for cropped objects): Given an input image and its pixel-aligned point cloud, we crop both using the visible instance mask. The image is cropped to a square bounding box around the mask, enlarged by a factor of 1.2, and background pixels are set to black. The points are centered by subtracting their centroid and scaled by twice the maximum absolute coordinate of the centered cloud, so that the observed points lie within the unit cube [-0.5,0.5]3. The network makes its predictions in this normalized frame, and we map them back to the input frame by inverting the scaling and translation. Because the network outputs explicit superquadric parameters rather than decoding over a fixed volume, nothing constrains the predicted primitives to lie within this cube. This allows the model to complete geometry that extends beyond the observed points.

Further, the reviewer is curious about how the training on ShapeNet can provide robustness to occlusions. Training on single objects from ShapeNet, would only provide robustness to self-occlusions, but not to occlusions coming from image cropping and object-to-object occlusion. However, as we explain in the paper (page 6, around lines 307–309), our model is not only trained on ShapeNet, but also fine-tuned on ASE and TableVerse, which are scene level datasets which naturally contain object-object occlusions. Training first on object level dataset and then on occluded ones is a common recipe for single-view reconstruction methods also used in SAM3D and ShapeR. We will explain this point better in the revised paper.

To verify that this robustness generalizes beyond the fine-tuning data, we additionally evaluate all methods zero-shot on two real-world datasets that are not part of any method’s training data: Aria Digital Twin (ADT) and HouseCat6D (HC), the latter featuring cluttered tabletop scenes with frequent inter-object occlusions. PixelDec outperforms both baselines in IoU on both datasets (0.405 vs. 0.369 on ADT and 0.683 vs. 0.653 on HC, compared to the strongest baseline, ShapeR + SuperFlex), confirming that its robustness to occlusion is not limited to the datasets it was fine-tuned on (see our response to reviewer z4Th for full results). We will clarify this in the revised paper.

3. Performance on Real-World Scenes The reviewer questions the robustness of our method in real-world indoor scenes, referring to the qualitative results in Figure 5. In particular, they highlight the bookshelf behind the sofa, and the drawer behind the column. We agree that our method is not perfect and does not work 100% of the time, but these are generally hard cases and also baselines struggle similarly. For these particular objects, we additionally include the results of the baseline methods in Fig. 15 (last two columns), while we report full results on those scenes in Fig. 21. We reiterate that our main objective is not perfect 3D reconstruction but a geometric approximation that is obtained at a fraction of the computational cost (>300x faster, ~10× less GPU memory). PixelDec reconstructs entire scenes in less than a second, while ShapeR and SAM3D need several minutes. We report additional details about our efficiency in Fig. 16.

We also added Fig. 19 and Fig. 20 which provide more insights on how different levels of occlusions affect different methods.

4. Performance on Robotics Scenes Regarding Fig. 9 in the paper, the reviewer asks why the handle of the mug and the pot are not reconstructed. For the mug handle, the mug is rotated such that the handle is not visible from the camera view-point (the camera is mounted between the two arms). For the pot, the handles are thin and reflective, so they are not well reconstructed in the depth of the stereo camera. We show the camera image and reconstructed depth in Fig. 17 of the appendix in the updated paper. In general however, our method is able to reconstruct handles, as illustrated in Fig. 13.

I thank the authors for the response, but several key concerns remain.

1. The completion ability is not well explained (W1). In Fig. 20 (first column, 12% observed), nothing in the input indicates that the object is a sofa. The model can only guess, and a conservative output such as SAM3D's seems more reasonable here. The rebuttal does not explain what in the model supplies such a strong category prior. Together with the weaker results on real scenes and the incomplete robot reconstructions, this points to in-domain overfitting rather than real robustness to occlusion.

2. The new zero-shot experiments (Tab. 11) raise further concerns. The zero-shot results on ADT and HouseCat6D are informative, but they also raise new questions. The experimental setting is not described. For example, it is unclear whether the inputs are GT depth and masks or predicted point maps and masks, and how many objects are evaluated. The "‡" footnote on PixelDec is never explained, and Tab. 11 is not referenced anywhere in the supplementary material. I would appreciate it if the authors could clarify these points and, ideally, provide qualitative comparisons under the Tab. 11 setting.

3. The contribution is unclear. In the robot experiment, the pot handles are not occluded and are clearly visible in the RGB input, yet PixelDec still fails to reconstruct them. This is a simple case that does not even involve occlusion. If the contribution is low-cost modeling, the main paper should evaluate on general scenes rather than focusing on heavily occluded cases. If it is occlusion handling, failing in such simple scenes undermines that claim. I would appreciate it if the authors could clarify the intended positioning of the paper, as this would help me better assess its contribution.

Overall, I will maintain my rating of weak reject.

Reviewer_z4ThOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of recovering a complete, pixel-aligned superquadric decomposition of every object in a scene from a single RGB image. Existing approaches either fit superquadrics to point clouds with optimization or feed-forward networks, or first generate a dense mesh with an image-to-3D model and then fit primitives to it. The former only explain observed geometry or regress a single deterministic completion that averages over ambiguous unseen parts, while the latter are slow, memory-heavy and spend most of their compute on a dense intermediate that is discarded. The authors instead treat superquadric parameters as the state space of a flow-matching diffusion transformer, conditioned on DINOv2 image features and a pixel-aligned point map, trained by distilling a SuperFlex teacher run on complete point clouds, with continuous flow for geometry and discrete flow for slot existence. This yields complete decompositions of all objects in a scene in a single batched pass under one second, and the authors validate it against primitive decomposition baselines on ShapeNet, against generate-then-fit pipelines on two synthetic scene datasets, and through simulated and real-robot grasping and a VLM-driven planning demo that consume the primitives directly.

Critical strengths
  1. The core idea of running flow matching directly in superquadric parameter space is simple and well motivated: it avoids the dense intermediate of generate-then-fit pipelines and models the ambiguity of unobserved geometry that deterministic regressors average away, which the ShapeNet results and the qualitative comparisons with SuperFlex support convincingly.

  2. The resulting representation is shown to be useful downstream, not only as an output format: the decluttering experiment demonstrates a clear grasp success improvement over SuperFlex primitives on the same planners, and the sub-second scene decomposition makes closed-loop use realistic, which is the main practical argument for primitive-based scene representations.

Critical weaknesses
  1. The comparison on ASE and TableVerse is not apples to apples: PixelDec is fine-tuned on both datasets for hundreds of epochs, whereas ShapeR and SAM3D are applied zero-shot (SAM3D additionally through a rigid fit), so the reported gains partly reflect in-domain adaptation rather than the method itself.

  2. Fidelity to the observed geometry is weaker than the baselines but not discussed: on ASE the point-to-surface distance of PixelDec (2.55) is clearly worse than ShapeR (1.11) and ShapeR+SuperFlex (1.75), which matters for manipulation where the visible side is exactly the part that must be accurate.

  3. The distillation setup has an unreported ceiling: the targets are SuperFlex outputs on complete point clouds, yet the teacher's own accuracy on complete clouds is never given, and the fixed slot ordering inherited from the teacher is assumed consistent across objects without any analysis.

  4. The robotics evidence is thin relative to the claims: the simulated decluttering uses ground-truth instance masks, the real-robot deployment reports no quantitative results, and the VLM agent experiment is a single qualitative demo without a comparison against an agent operating on point clouds or meshes.

Minor: the generation metrics are underspecified (how points are sampled from primitives for P-FID/P-KID, the undefined Diversity metric in Tab. 4, negative P-KID values), the choice of 4 inference steps optimizes IoU while P-FID and P-KID keep improving at more steps, and there are several writing issues (incomplete sentence "SuperFlex for our same task", "boxs", "tableverse").

Questions for authors

Questions:

  1. Are ShapeR and SAM3D evaluated zero-shot in Tab. 2 while PixelDec is fine-tuned on ASE and TableVerse? If so, please report PixelDec without fine-tuning or the baselines with comparable adaptation.

  2. What is the accuracy of the SuperFlex teacher on complete point clouds, so the gap between the student and its supervision ceiling can be assessed?

  3. Why is the point-to-surface distance on ASE worse than the baselines, and does this affect grasps on the visible side of objects?

  4. How is the highest-likelihood sample selected among the 10 draws in the robotics experiments, given that the flow model has no explicit likelihood?

Suggestion: a quantitative evaluation on real captured data with ground-truth meshes would strengthen the generalization claim.

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback and the discussions throughout the rebuttal. Their questions motivated substantial additional experiments and analyses, which have strengthened the paper and provided further evidence supporting our contributions. We would like to summarize how these additions address the main concerns raised in the original reviews.

  • Generalization and fairness of evaluation (z4Th, D2bc). We added zero-shot quantitative and qualitative evaluations on two real-world datasets (Aria Digital Twin and HouseCat6D), including experiments with predicted instance masks (SAM3) and point maps (MoGe-2). PixelDec achieves competitive or superior reconstruction quality without in-domain fine-tuning, while remaining over 1,000× faster than the baselines in scene-level reconstruction (Figs. 23–24).
  • Benefits of generative modeling (pdB3). We implemented and trained a deterministic version of PixelDec with matched architecture, conditioning, and supervision. Quantitative and qualitative comparisons (Tab. 9 and Fig. 18) demonstrate that the substantial improvements stem primarily from generative modeling rather than simply incorporating image features.
  • Robustness to occlusions (D2bc, pdB3). We conducted additional quantitative (Fig. 19) and qualitative (Fig. 20) evaluations across different levels of visibility and provided detailed analyses of the failure modes of SAM3D and ShapeR. As discussed in our rebuttal responses, these analyses explain why PixelDec better handles partial observations than the baselines.
  • Comparison with prior work (pdB3, D2bc). We provided detailed comparisons with previous deterministic superquadric reconstruction methods (SuperDec, SuperFlex), highlighting the benefits of our generative formulation for completing partially observed geometry. We also clarified the differences from concurrent work DoSs, which focuses on unconditional superquadric generation for a limited set of object categories in canonical poses. In contrast, PixelDec performs class-agnostic, observation-conditioned generation directly in camera coordinates, using deformable superquadrics and a flow-matching formulation tailored to the different types of superquadric parameters.
  • Computational efficiency (pdB3, D2bc). We expanded the analysis to include end-to-end latency, preprocessing, GPU memory usage, and runtime scaling with scene complexity. PixelDec reconstructs entire scenes in under a second, compared with several minutes for the baselines, demonstrating the practical benefits of generating directly in a compact representation.
  • Practical utility for robotics (z4Th, D2bc, pdB3). We quantitatively evaluated the impact of predicted instance masks on robotic grasping (Tab. 8), demonstrating that PixelDec maintains the highest grasp success rates even under realistic segmentation conditions. These results further demonstrate that, despite imperfect geometric reconstructions, our representations remain effective and provide measurable benefits for downstream robotic tasks.

We believe these additional experiments address the majority of the original concerns and provide substantially stronger evidence for the two main advantages of our approach: significantly greater efficiency than state-of-the-art image-to-3D generative methods (SAM3D, ShapeR), and substantially improved reconstruction quality over state-of-the-art deterministic superquadric decomposition methods (SuperDec, SuperFlex). We would also like to reiterate that our primary contribution is generative 3D scene decomposition, with robotics serving as a downstream demonstration of the representation's practical utility.

We hope that the substantial additional results and clarifications provided throughout the rebuttal will be fully considered in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.