acceptodds
Under review as a conference paper at ICLR 2027

ScaleBlind: Point Cloud Completion under Unknown Scale

Abstract

Point cloud completion aims to infer a complete 3D shape from a partial point cloud and serves as a fundamental building block for downstream tasks such as reconstruction, editing, and simulation. Despite the recent progress, existing learning-based methods often implicitly rely on access to the ground-truth shape scale (GT-scale) during both training- and testing-time normalization, assuming privileged information that is unavailable in real-world inference. This hidden assumption limits practical deployment and can lead to severe completion artifacts, e.g., over- or under-completion and nested shells, once the oracle GT-scale cue is removed. We observe that the recent foundation image generation models exhibit a strong capability of understanding objects and geometries, and producing multi-view consistent renderings, making them promising priors for GT-scale-free 3D completion. Motivated by this insight, we propose ScaleBlind, a novel framework that leverages foundation-model-based image completion to recover global scale directly from partial inputs and then faithfully produces the 3D completion. Specifically, ScaleBlind dreams out complete multi-view appearances from rendered partial views, lifts the inferred missing regions back into 3D to obtain a geometry-aware coarse completion, and further refines it via a powerful cross-modal fusion network with the original partial point cloud. By harnessing 2D foundation priors, our method eliminates the need for accessing GT-scale information at inference. Moreover, it provides a principled bridge between 2D generative priors and 3D point cloud completion. Extensive experiments demonstrate the superiority of our framework, making ScaleBlind the new state-of-the-art for the point cloud completion task.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point Cloud Completion, Diffusion, Multi-modality Fusing

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_PZy4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ScaleBlind, a point cloud completion framework designed for scenarios where the ground-truth shape scale is unavailable during inference. The authors highlight that conventional point cloud completion benchmarks often rely on ground-truth-based normalization, which introduces privileged scale information that may not be accessible in practical applications. To address this issue, ScaleBlind leverages a pretrained image generation model to complete multi-view renderings of partial point clouds, reconstructs a coarse geometric prior through voxel voting, and incorporates the recovered prior into a coarse-to-fine completion network. Experiments demonstrate promising completion performance under both GT-scale-free and conventional evaluation settings.

Critical strengths
  1. The paper identifies an important limitation of conventional point cloud completion protocols and investigates completion under unknown global scale. This setting is practically relevant, particularly when substantial missing regions prevent partial observations from preserving the full spatial extent of an object.

  2. The proposed framework provides a coherent solution by combining foundation-model-based image completion, multi-view geometric prior reconstruction, and prior-guided point cloud refinement. The experimental results demonstrate competitive performance and improvements over several existing methods, supporting the potential of the proposed approach.

Critical weaknesses
  1. The experimental comparisons omit several recent point cloud completion methods, such as SymmCompletion (AAAI'25) and Hyper-PCN (CVPR'26). More importantly, under the GT-scale-free setting, most existing baselines are provided with scale estimates produced by ScaleBlind's prior acquisition module. While this is a reasonable strategy for comparing completion architectures under a shared scale estimate, it does not fully establish that conventional completion architectures are inherently unable to handle unknown-scale inputs. A more informative comparison would retrain representative baselines, such as AdaPoinTr, using partial-only normalization during both training and inference. This would help clarify whether the reported improvements primarily originate from the proposed scale estimation mechanism, the completion architecture, or their combination.

  2. Recovering the unknown global extent of incomplete objects is a central motivation and contribution of this paper. However, the current evaluation primarily relies on conventional completion metrics, including CD, DCD, and F-Score. These metrics assess the quality of the final completed point clouds but do not directly measure the accuracy of the recovered scale. I suggest reporting explicit scale estimation metrics, such as the relative error between the estimated and ground-truth shape scales.

  3. The current ablation studies mainly investigate the confidence attribute, hierarchical image features, and the number of viewpoints. Although these results are useful, they do not sufficiently disentangle the contributions of the major components of the proposed framework. In particular, it would be helpful to evaluate the completion network without the generated geometric prior, with scale estimation alone, and with different combinations of geometric candidate points and image features. Additionally, the computational analysis could be more comprehensive. Since the framework relies on multiple image generation calls and volumetric prior reconstruction, its practical efficiency is an important consideration. Although the supplementary material includes an analysis of inference time with different numbers of viewpoints, a breakdown of the complete inference pipeline, including image generation, voxel voting, and point cloud refinement, would better characterize the computational overhead.

Questions for authors
  1. Have the authors considered retraining conventional completion methods under partial-only normalization, without using ScaleBlind's estimated scale? How would their performance compare with the proposed method?

  2. Can the authors provide quantitative results for the recovered scale itself, rather than evaluating only the final completion quality? How does scale estimation accuracy vary with the missing ratio?

  3. What is the relative contribution of the scale estimation module and the prior-guided completion network? In particular, how would the performance change if the estimated scale were retained but the generated geometric prior and image features were removed?

  4. What is the end-to-end inference time per object, including the multi-view image generation process? How much computational overhead does the proposed framework introduce compared with conventional point cloud completion methods?

6 comments

Thank you for your insightful comments. We answer your concerns and questions in the following.

W1-1: The experimental comparisons omit several recent point cloud completion methods, such as SymmCompletion (AAAI'25) and Hyper-PCN (CVPR'26).

A: We additionally evaluate SymmCompletion (AAAI 2025) and Hyper-PCN (CVPR 2026) under the same GT-scale-free and GT-scale-tied protocols as Table 1. Under the GT-scale-free setting, both baselines use our estimated scales for normalization, consistent with the original comparisons. The additional baseline results have been incorporated into Sec. 3.1 (Table 1) of the revised paper.

The results are reported below, with each entry corresponding to GT-scale-free / GT-scale-tied.

MethodsPlaneCabinetCarChairLampCouchTableBoatCD-Avg ↓DCD-Avg ↓F1 ↑
SymmCompletion4.13 / 3.917.44 / 7.127.59 / 6.876.81 / 6.398.98 / 7.796.87 / 5.976.33 / 5.516.52 / 5.716.83 / 6.160.525 / 0.5110.837 / 0.864
Hyper-PCN4.22 / 3.617.17 / 6.637.36 / 6.946.78 / 6.318.83 / 7.306.91 / 6.336.17 / 5.596.64 / 5.836.76 / 6.070.521 / 0.5050.843 / 0.875
ScaleBlind (Ours)3.88 / 3.567.39 / 7.037.08 / 6.846.49 / 6.097.34 / 7.106.18 / 5.875.72 / 5.445.89 / 5.666.21 / 5.950.514 / 0.4990.871 / 0.883

ScaleBlind achieves the best overall performance across all three metrics under both protocols. Under GT-scale-free evaluation, its CD-Avg is 6.21, compared with 6.76 for Hyper-PCN and 6.83 for SymmCompletion.

W1-2: More importantly, under the GT-scale-free setting, most existing baselines are provided with scale estimates produced by ScaleBlind's prior acquisition module. While this is a reasonable strategy for comparing completion architectures under a shared scale estimate, it does not fully establish that conventional completion architectures are inherently unable to handle unknown-scale inputs. A more informative comparison would retrain representative baselines, such as AdaPoinTr, using partial-only normalization during both training and inference. This would help clarify whether the reported improvements primarily originate from the proposed scale estimation mechanism, the completion architecture, or their combination.

A: We additionally retrain five representative baselines (AdaPoinTr, SVDFormer, CRA-PCN, GeoFormer, and PCDreamer) from scratch using partial-only normalization during both training and testing, eliminating the train–test normalization mismatch. We compare these results with the GT-scale-free results in Table 1, where the baselines use our estimated scales, to distinguish the benefits of scale information and the completion framework.

For each baseline, the two values correspond to Partial-only-Retrained) / With-Our-Scale (Table 1), respectively. ScaleBlind is evaluated under its original GT-scale-free setting. We only report the average CD, DCD, and F1 in the following, the full numbers can be found in Appendix F.7 (Table F4).

MethodsCD-Avg ↓DCD-Avg ↓F1 ↑
AdaPoinTr10.85 / 7.070.615 / 0.5300.691 / 0.826
SVDFormer10.95 / 7.180.616 / 0.5320.704 / 0.818
CRA-PCN14.80 / 6.940.722 / 0.5230.502 / 0.834
GeoFormer14.24 / 7.250.656 / 0.5350.582 / 0.814
PCDreamer9.73 / 7.050.581 / 0.5280.736 / 0.829
ScaleBlind (Ours)6.210.5140.871

Result interpretation: Comparing ScaleBlind with baselines using our estimated scales (With-Our-Scale) reveals the advantage of our prior-guided completion framework under shared scale information. Comparing ScaleBlind with baselines retrained under partial-only normalization (Partial-only-Rerained) reflects the combined benefits of scale estimation and completion. The difference between Partial-only-Retrained and With-Our-Scale further indicates the benefit that our estimated scales provide to conventional completion methods.

In general, across all three metrics, ScaleBlind obtains the best performance against all baselines. Particularly, the gain comes primarily from the scale estimation mechanism.

Take PCDreamer as an example, ScaleBlind reduces CD-Avg from 7.05 to 6.21 (11.9% reduction), demonstrating the advantage of our prior-guided completion framework. ScaleBlind achieves a larger CD reduction from 9.73 to 6.21 (36.2% reduction) compared with the Partial-only-Retrained version, reflecting the combined benefits of scale estimation and prior-guided completion. The difference between these two improvements (3.52 − 0.84 = 2.68) provides an approximate indication of the benefit associated with our estimated scale. These results highlight the complementary contributions of scale recovery and the completion framework.

W2: Recovering the unknown global extent of incomplete objects is a central motivation and contribution of this paper. However, the current evaluation primarily relies on conventional completion metrics, including CD, DCD, and F-Score. These metrics assess the quality of the final completed point clouds but do not directly measure the accuracy of the recovered scale. I suggest reporting explicit scale estimation metrics, such as the relative error between the estimated and ground-truth shape scales.

A: We introduce Global Scale Error (GSE) to directly measure the relative error between the estimated global extent and the ground-truth complete shape. The detailed formulation and evaluation procedure are provided in Appendix E.7 (Eqs. (8)–(10) and Table E5) of the revised paper.

Specifically, during data construction, we divide each complete shape into the partial observation and the removed region. We independently normalize the partial observation and generate geometric candidates through multi-view voxel voting. To evaluate scale recovery, we align the normalized partial observation back to its original coordinates and apply the same inverse transformation to the voting candidates. We then combine the transformed candidates with the original partial observation and compute GSE as the relative difference between the bounding-box diagonal lengths of the estimated complete shape and the GT. GT information is used only for evaluation, not for constructing the geometric priors.

The quantitative results under different missing ratios are reported below.

Missing RatioGSE ↓
25%7.04%
50%9.31%
75%12.52%

GSE increases from 7.04% to 12.52% as the missing ratio rises from 25% to 75%, indicating that scale estimation becomes more challenging with greater incompleteness. These results directly quantify the accuracy of our recovered global extent, with a relative error of 12.52% even at a 75% missing ratio.

W3-1: The current ablation studies mainly investigate the confidence attribute, hierarchical image features, and the number of viewpoints. Although these results are useful, they do not sufficiently disentangle the contributions of the major components of the proposed framework. In particular, it would be helpful to evaluate the completion network without the generated geometric prior, with scale estimation alone, and with different combinations of geometric candidate points and image features.

A: We conduct additional ablation experiments to separately examine the contributions of scale estimation alone, geometric candidate points, and image features, as well as the effects of geometric prior confidence and the number of image views used for completion. These experiments complement the existing ablations on confidence attributes and hierarchical image features in Table 2 and the viewpoint analysis in Fig. E3. These additional experiments are included in Appendices E.4–E.6 (Tables E2–E4) of the revised paper.

1. Scale Estimation and Component Ablation

To directly assess the contribution of scale estimation alone, we retain the estimated scale obtained from P_merge for normalization, but remove both the voting candidates and image features from the completion network. We also evaluate two additional configurations: (1) w/o Joint Normalization, where we use the independently normalized partial point cloud as the geometric input while retaining image features; and (2) w/o Image Features, where we retain P_merge-based normalization and geometric candidates but remove the image feature branch.

MetricFull Modelw/o Joint NormalizationScale Estimation Onlyw/o Image Features
CD-Avg ↓6.218.897.817.52
DCD-Avg ↓0.5140.5760.5630.559
F1 ↑0.8710.7640.7930.801

The scale-estimation-only variant achieves a CD-Avg of 7.81, demonstrating that the estimated scale can support completion even without geometric candidates or image features as network inputs. Adding geometric candidates further reduces CD to 7.52, while incorporating image features improves it to 6.21. Similar improvements are observed in DCD and F1. In comparison, using only the independently normalized partial point cloud with image features yields a CD of 8.89. These results highlight the complementary benefits of estimated scale, geometric candidates, and image features in our framework.

2. Generated 3D Priors (Confidence Threshold)

To investigate the influence of generated geometric priors, we vary the confidence threshold τ ∈ {0.0, 0.2, 0.4, 0.6, 0.8, 1.0}, while keeping the partial input and 13 image features unchanged. At τ = 0, we retain voxels with probability > 0 to exclude zero-probability voxels. For all other thresholds, we retain voxels with probability ≥ τ.

Metricτ = 0.0τ = 0.2τ = 0.4τ = 0.6τ = 0.8 (Default)τ = 1.0
CD-Avg ↓11.738.977.176.396.216.71
DCD-Avg ↓0.7280.5830.5510.5280.5140.542
F1 ↑0.6210.7520.8120.8630.8710.850

Lower thresholds retain more unreliable geometric candidates and degrade completion performance, while overly strict thresholds may discard useful geometric information. The default threshold τ = 0.8 achieves the best performance across all three metrics, confirming the importance of confidence-aware geometric prior selection.

3. Image Features (Number of Views)

To isolate the contribution of image features, we fix the geometric prior and estimated scale for each sample, varying only the number of images provided to the completion network. Specifically, we randomly select 1, 3, 5, 7, 9, or 11 views from the original 13, and train each setting independently using reproducible random sampling.

Metric1 View3 Views5 Views7 Views9 Views11 Views13 Views (Default)
CD-Avg ↓6.686.606.546.456.376.296.21
DCD-Avg ↓0.5410.5370.5360.5310.5270.5200.514
F1 ↑0.8520.8550.8570.8620.8640.8670.871

With geometric priors and estimated scales fixed, increasing the number of image views consistently improves completion performance. CD-Avg decreases from 6.68 with one view to 6.21 with 13 views, demonstrating that multi-view image features provide complementary guidance for geometric refinement.

Reviewer_qageOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ScaleBlind, a point cloud completion network addressing the unknown global scale problem. Previous completion methods usually rely on access to the ground-truth shape scale, but struggle when this information is unknown, limiting their applicability to real-world scenarios. To address this problem, ScaleBlind uses a pretrained VLM/image generator to inpaint missing regions, lift the completed 2D evidence back to 3D through a visual-hull-like voting scheme, and then refine the result with a coarse-to-fine point completion network.

Critical strengths
  • The paper addresses a clear, well-defined problem that has received limited attention in prior work.
  • The use of a VLM for multi-view image inpainting is an interesting and novel idea.
  • Overall, the paper is clearly written and easy to follow.
Critical weaknesses
  • The ablations do not clearly separate the effects of joint normalization, the generated 3D prior, and image features. More targeted ablations and direct scale-error measurements would better support the central contribution.
  • The source of the test-time priors needs clarification. Appendix D.1 describes GT-rendered completed views and mentions their use for “training and testing,” whereas Fig. D1 states that simulated priors are used only during training. Please confirm that all quantitative test results use priors generated from the partial observations without access to the complete GT geometry.
Questions for authors

See the weaknesses.

4 comments

Thank you for your insightful comments. We answer your concerns and questions in the following.

W1-1: The ablations do not clearly separate the effects of joint normalization, the generated 3D prior, and image features.

A: We conduct additional ablation experiments to separately investigate the contributions of prior-based joint normalization, generated 3D geometric priors, and image features. These experiments complement our existing ablations on confidence attributes and hierarchical image features in Table 2, as well as the viewpoint analysis in Fig. E3. The three additional ablation studies are presented in Appendices E.4–E.6 (Tables E2–E4) of the revised paper.

1. Scale Estimation and Component Ablation

To directly assess the contribution of scale estimation alone, we retain the estimated scale obtained from P_merge for normalization, but remove both the voting candidates and image features from the completion network. We also evaluate two additional configurations: (1) w/o Joint Normalization, where we use the independently normalized partial point cloud as the geometric input while retaining image features; and (2) w/o Image Features, where we retain P_merge-based normalization and geometric candidates but remove the image feature branch.

MetricFull Modelw/o Joint NormalizationScale Estimation Onlyw/o Image Features
CD-Avg ↓6.218.897.817.52
DCD-Avg ↓0.5140.5760.5630.559
F1 ↑0.8710.7640.7930.801

The scale-estimation-only variant achieves a CD-Avg of 7.81, demonstrating that the estimated scale can support completion even without geometric candidates or image features as network inputs. Adding geometric candidates further reduces CD to 7.52, while incorporating image features improves it to 6.21. Similar improvements are observed in DCD and F1. In comparison, using only the independently normalized partial point cloud with image features yields a CD of 8.89. These results highlight the complementary benefits of estimated scale, geometric candidates, and image features in our framework.

2. Generated 3D Priors (Confidence Threshold)

To investigate the influence of generated geometric priors, we vary the confidence threshold τ ∈ {0.0, 0.2, 0.4, 0.6, 0.8, 1.0}, while keeping the partial input and 13 image features unchanged. At τ = 0, we retain voxels with probability > 0 to exclude zero-probability voxels. For all other thresholds, we retain voxels with probability ≥ τ.

Metricτ = 0.0τ = 0.2τ = 0.4τ = 0.6τ = 0.8 (Default)τ = 1.0
CD-Avg ↓11.738.977.176.396.216.71
DCD-Avg ↓0.7280.5830.5510.5280.5140.542
F1 ↑0.6210.7520.8120.8630.8710.850

Lower thresholds retain more unreliable geometric candidates and degrade completion performance, while overly strict thresholds may discard useful geometric information. The default threshold τ = 0.8 achieves the best performance across all three metrics, confirming the importance of confidence-aware geometric prior selection.

3. Image Features (Number of Views)

To isolate the contribution of image features, we fix the geometric prior and estimated scale for each sample, varying only the number of images provided to the completion network. Specifically, we randomly select 1, 3, 5, 7, 9, or 11 views from the original 13, and train each setting independently using reproducible random sampling.

Metric1 View3 Views5 Views7 Views9 Views11 Views13 Views (Default)
CD-Avg ↓6.686.606.546.456.376.296.21
DCD-Avg ↓0.5410.5370.5360.5310.5270.5200.514
F1 ↑0.8520.8550.8570.8620.8640.8670.871

With geometric priors and estimated scales fixed, increasing the number of image views consistently improves completion performance. CD-Avg decreases from 6.68 with one view to 6.21 with 13 views, demonstrating that multi-view image features provide complementary guidance for geometric refinement.

W1-2: More targeted ablations and direct scale-error measurements would better support the central contribution.

A: We introduce Global Scale Error (GSE) to directly measure the relative error between the estimated global extent and the ground-truth complete shape. The detailed formulation and evaluation procedure are provided in Appendix E.7 (Eqs. (8)–(10) and Table E5) of the revised paper.

Specifically, during data construction, we divide each complete shape into the partial observation and the removed region. We independently normalize the partial observation and generate geometric candidates through multi-view voxel voting. To evaluate scale recovery, we align the normalized partial observation back to its original coordinates and apply the same inverse transformation to the voting candidates. We then combine the transformed candidates with the original partial observation and compute GSE as the relative difference between the bounding-box diagonal lengths of the estimated complete shape and the GT. GT information is used only for evaluation, not for constructing the geometric priors.

The quantitative results under different missing ratios are reported below.

Missing RatioGSE ↓
25%7.04%
50%9.31%
75%12.52%

GSE increases from 7.04% to 12.52% as the missing ratio rises from 25% to 75%, indicating that scale estimation becomes more challenging with greater incompleteness. These results directly quantify the accuracy of our recovered global extent, with a relative error of 12.52% even at a 75% missing ratio.

W2: The source of the test-time priors needs clarification. Appendix D.1 describes GT-rendered completed views and mentions their use for “training and testing,” whereas Fig. D1 states that simulated priors are used only during training. Please confirm that all quantitative test results use priors generated from the partial observations without access to the complete GT geometry.

A: All test-time geometric priors are generated solely from partial observations through VLM-based image completion and voxel voting, without accessing complete GT geometry. GT-rendered simulated priors are used exclusively for training data preparation. Under the GT-scale-free setting, no GT-scale information is accessed during inference.

We apologize for the inaccurate wording in Appendix D.1 and have corrected the relevant description in the manuscript (line 821) to clarify this distinction.

My concerns have been addressed and I will raise my score.

We are grateful for your encouraging feedback and continued support of our work. Your insightful comments motivated us to conduct more targeted ablation studies, directly evaluate scale estimation accuracy, and clarify the use of geometric priors during inference. These improvements have made our experimental evidence more comprehensive and the paper more rigorous. We sincerely appreciate your careful review and valuable suggestions.

Reviewer_J2QbOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates point cloud completion when the scale of the complete object is unknown. Common benchmarks first normalize complete shapes and then generate partial inputs, potentially introducing implicit priors about the scale and position of the complete shape into the input. The absence of this information during real-world inference can lead to distorted completions. To address this issue, ScaleBlind first uses a pretrained image generation model to complete multi-view renderings of the partial point cloud, then constructs a 3D prior for the missing regions with associated confidence scores through voxel voting. This prior is used to estimate the spatial extent of the complete object and perform joint normalization. It also guides a coarse-to-fine completion network that integrates the point cloud with hierarchical image features.

Under the unknown-scale setting on PCN, the average CD decreases and the gains are smaller when the true object scale is known. Thus the experimental results support the effectiveness of the proposed method.

Critical strengths
  1. The paper clearly distinguishes between two settings: one in which the input is already expressed in the normalized coordinate system of the complete object, and another in which scale must be inferred solely from partial observations. It also demonstrates the distortions that can arise when the former assumption is removed. This issue directly affects whether results obtained on synthetic benchmarks can transfer to real-world deployment, and addressing it is the paper’s most valuable contribution.

  2. Generated images are not used merely as auxiliary features. They are converted into a 3D confidence prior for spatial extent estimation, normalization, and geometric refinement. Compared with directly using an image-to-3D generator, this design places greater emphasis on alignment with the observed geometry. The overall improvements across experiments on PCN, ShapeNet-55, and single-view scans provide reasonably consistent evidence for the framework’s effectiveness.

Critical weaknesses

The failure cases under partial-only normalization primarily involve models trained with GT-scale normalization, thus conflating the effects of missing scale information with those of distribution shift. Without strong baselines trained from scratch under the partial-only protocol, the paper does not sufficiently establish that its complex VLM-based scale recovery pipeline outperforms simpler alternatives, such as adapting the normalization protocol or directly predicting scale.

Questions for authors

How would strong baselines perform in the unknown-scale setting if partial-only normalization were adopted from the start of training and appropriate scale augmentation were applied? Could you provide quantitative results for at least one strong baseline under this matched protocol to distinguish the effects of missing scale information itself from those of train–test distribution shift?

6 comments

Thank you for your insightful comments. We answer your concerns and questions in the following.

W1: The failure cases under partial-only normalization primarily involve models trained with GT-scale normalization, thus conflating the effects of missing scale information with those of distribution shift. Without strong baselines trained from scratch under the partial-only protocol, the paper does not sufficiently establish that its complex VLM-based scale recovery pipeline outperforms simpler alternatives, such as adapting the normalization protocol or directly predicting scale.

A: We additionally retrain five representative baselines (AdaPoinTr, SVDFormer, CRA-PCN, GeoFormer, and PCDreamer) from scratch using partial-only normalization during both training and testing, thereby eliminating the train–test normalization mismatch. Unlike the GT-scale-free comparisons in Table 1, these baselines no longer use our estimated scales, allowing us to directly evaluate whether conventional completion networks can handle unknown-scale inputs through partial-only normalization alone. The detailed results of this additional evaluation have been incorporated into Appendix F.7 (Table F4) of the revised paper.

The quantitative results on PCN are reported below:

MethodsPlaneCabinetCarChairLampCouchTableBoatCD-Avg ↓DCD-Avg ↓F1 ↑
AdaPoinTr6.2410.109.3910.5219.739.2210.6810.9310.850.6150.691
SVDFormer6.129.869.4410.7222.068.829.9810.6310.950.6160.704
CRA-PCN8.5715.3212.1915.4124.9913.9514.2213.7214.800.7220.502
GeoFormer8.5313.4612.6814.5724.2913.7113.9112.7314.240.6560.582
PCDreamer5.669.208.619.7516.308.549.889.909.730.5810.736
ScaleBlind (Ours)3.887.397.086.497.346.185.725.896.210.5140.871

ScaleBlind achieves the best performance across all three metrics, reducing the average CD from 9.73 to 6.21 (36.2% reduction) compared with the strongest retrained baseline, PCDreamer, which reflects the effectiveness of our scale-estimation scheme.

Additionally, compared with the GT-scale-free results reported in Table 1, where the baselines are evaluated using our estimated scales, all five baselines exhibit substantially worse performance after being retrained with partial-only normalization. This indicates the difficulty of the practical unknown-scale setting.

Thank you for your response. In my view, this work addresses an important problem in point cloud completion. However, establishing a fair comparison with existing methods remains challenging, as some approaches inherently rely on ground-truth normalization and may not function properly when this assumption is removed. Furthermore, regarding computational efficiency, I noticed that the VLM accounts for a substantial portion of the inference time, yet no comparison with a simpler alternative is provided. Given the additional computational cost and the scale-related challenges considered in this work, I believe the authors should further justify the necessity of introducing a VLM and clarify whether its contribution warrants the associated overhead.

We appreciate the reviewer's continued engagement and recognition of the importance of the problem addressed in our work. We further clarify the fairness of our comparisons and the role of the VLM below.

Q1: However, establishing a fair comparison with existing methods remains challenging, as some approaches inherently rely on ground-truth normalization and may not function properly when this assumption is removed.

A: We agree that establishing a fully equivalent comparison with existing point cloud completion methods is challenging, since these methods were not originally designed to address the unknown global scale problem. To ensure a fair comparison to the greatest extent possible, we have carefully designed and conducted evaluations under three complementary protocols:

1. GT-scale-tied (Oracle Scale): The methods are evaluated using GT-scale information for normalization, following the conventional protocol used in their methods. This ensures their inherent dependence on GT normalization and keeps all their benefits in their method design. However, this setting is impractical and provides a performance upper-bounds.

2. GT-scale-free (Our Estimated Scale): We provide our estimated scales to the applicable baselines for normalization. Because our prediction has reasonable errors (check the new scale metric and the discussion in Suppl. E7), it thus partially harms their inherent dependence on GT normalization (see table below).

3. Partial-only (Retrained, suggested by you): We additionally retrained five representative baselines (AdaPoinTr, SVDFormer, CRA-PCN, GeoFormer, and PCDreamer) from scratch using partial-only normalization during both training and testing. This eliminates the train–test normalization mismatch and allows these methods to adapt to unknown-scale inputs without relying on GT scales or our estimated scales. However, this potentially removes all of their inherent dependency on GT normalization, and proper adaptation of each method into the GT-scale-free setting is far beyond the scope of this paper.

The results under these three settings are summarized below (note, the ‘-’ means we did not evaluate the relevant methods under the corresponding setting):

MethodsCD-Avg (GT-scale-tied) ↓DCD-Avg (GT-scale-tied) ↓F1 (GT-scale-tied) ↑CD-Avg (GT-scale-free) ↓DCD-Avg (GT-scale-free) ↓F1 (GT-scale-free) ↑CD-Avg (Partial-only) ↓DCD-Avg (Partial-only) ↓F1 (Partial-only) ↑
AdaPoinTr6.460.5280.8527.070.5300.82610.850.6150.691
SVDFormer6.400.5100.8577.180.5320.81810.950.6160.704
CRA-PCN6.100.5070.8626.940.5230.83414.800.7220.502
GeoFormer6.290.5080.8697.250.5350.81414.240.6560.582
PCDreamer6.080.5030.8717.050.5280.8299.730.5810.736
SymmCompletion6.160.5110.8646.830.5250.837---
Hyper-PCN6.070.5050.8756.760.5210.843---
ScaleBlind (Ours)5.950.4990.8836.210.5140.871---

From the inter- and intra-setting comparison, we can find that:

  • ScaleBlind achieves the best overall performance across all three metrics under all three settings.
  • ScaleBlind under GT-scale-free setting vs. competitors under GT-scale-tied setting: ScaleBlind achieves a lower (better) CD-Avg (6.21) than AdaPoinTr (6.46), SVDFormer (6.40), and GeoFormer (6.29), while remaining close to the other four baselines: CRA-PCN (6.10), SymmCompletion (6.16), Hyper-PCN (6.07), and PCDreamer (6.08). Remember that the GT-scale information is more favorable to competitors, placing us in an unfair position. Even so, our performance is on par with theirs, but we have clear advantage of the real-world applicability.
  • Competitors under the three settings: as expected, all competitors get worse and worse results from GT-scale-tied, to GT-scale-free, and to Partial-only settings. For example, AdaPoinTr increases from 6.46 to 7.07 to 10.85, SVDFormer from 6.40 to 7.18 to 10.95. They fully exploit their inherent dependency on the GT-normalization in the GT-scale-tied setting, but it serves as the imagined performance upper bound. The later two settings are more practical, but their performance drops significantly.
Reviewer_L7LgOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ScaleBlind, which targets a long-overlooked global-scale leakage in point cloud completion: existing learned methods perform a center-and-scale normalization into the unit sphere that, in effect, uses the ground-truth complete shape at inference time — an "oracle" scale cue unavailable in deployment. The authors formalize a more deployment-realistic task, GT-scale-free point cloud completion, in which the global scale must be recovered from the occluded partial observation itself rather than assumed from GT.

Critical strengths

1.Identifying GT-scale-tied normalization as a hidden evaluation shortcut offers a reframing of an existing problem. 2.The proposed pipeline combines multi-view consistency, visual-hull voting, and prior-based joint normalization to recover a faithful scale without accessing ground truth, yielding a clean and self-consistent solution.

Critical weaknesses

1.The contribution is incremental: the method combines well-known modules and refines them, without a clearly new core mechanism.

2.Using a VLM for image completion can produce hallucinations and multi-view inconsistencies that affect the results, and the authors do not propose an effective means to address this problem.

3.The experiments on the datasets provided by the authors are too simple. The authors could consider testing on the Redwood dataset, such as the GenPc method from CVPR 2025.

4.The authors should provide comparisons of parameters and runtime. The authors mention in the experimental details that training uses 4 GPUs, which also raises concerns about training cost.

Questions for authors

1.The contribution is incremental: the method combines well-known modules and refines them, without a clearly new core mechanism.

2.Using a VLM for image completion can produce hallucinations and multi-view inconsistencies that affect the results, and the authors do not propose an effective means to address this problem.

3.The experiments on the datasets provided by the authors are too simple. The authors could consider testing on the Redwood dataset, such as the GenPc method from CVPR 2025.

4.The authors should provide comparisons of parameters and runtime. The authors mention in the experimental details that training uses 4 GPUs, which also raises concerns about training cost.

4 comments

Thank you for your insightful comments. We answer your concerns and questions in the following.

W1: The contribution is incremental: the method combines well-known modules and refines them, without a clearly new core mechanism.

A: Our contributions lie in identifying the overlooked GT-scale leakage problem and developing a dedicated framework for recovering unknown global scale from partial observations. The significance is also recognized by R2, who highlights the practical implications of GT-scale leakage; R3, who acknowledges the novelty of VLM-based multi-view inpainting; and R4, who recognizes the coherent design of our framework.

W2: Using a VLM for image completion can produce hallucinations and multi-view inconsistencies that affect the results, and the authors do not propose an effective means to address this problem.

A: This is a misunderstanding. Our framework mitigates VLM hallucinations and multi-view inconsistencies through geometric consistency filtering and confidence-aware voxel voting.

Specifically, we 1) reject unreliably generated views based on their foreground overlap with the observed partial projections; 2) aggregate evidence from the remaining views in a shared 3D voxel space, where consistently supported regions receive higher confidence. Inconsistent predictions can be suppressed through confidence-based filtering.

And, 3) the resulting geometric candidates serve as 3D priors rather than final completion results. They are subsequently refined by the completion network together with the observed partial point cloud and image features. Although this mechanism cannot entirely eliminate hallucinations that are consistent across multiple views, it effectively reduces the influence of view-specific generation errors on the reconstructed geometric prior.

W3: The experiments on the datasets provided by the authors are too simple. The authors could consider testing on the Redwood dataset, such as the GenPC method from CVPR 2025.

A: We additionally evaluate ScaleBlind on the Redwood dataset, using the same 13 objects with available GT references provided by GenPC (CVPR 2025). Since the original scans contain tens of thousands of points, we follow our paper's data construction protocol: removing 75% of points using KNN-based cropping, followed by farthest point sampling (FPS) to obtain 2,048 input points. Considering real-world deployment scenarios where complete GT geometry and its global scale are unavailable, we independently normalize the observed partial point clouds for all methods and run their respective complete inference pipelines without accessing GT-scale information. The available GT references are used exclusively for quantitative evaluation. The detailed experimental protocol and quantitative results have been incorporated into Appendix F.4 (Table F2) of the revised paper. The results are also reported below.

MethodWheelie BinPicnic TableChairArmchairPlant VasesTableVespaKid TricycleMotorcycleSofaTrash CanSwivel ChairBicycleCD-Avg ↓
SVDFormer26.2130.6826.8412.5950.8761.4941.7565.8526.7930.07131.86135.8149.9153.13
AdaPoinTr43.2633.9835.1917.0687.3358.5259.9673.1633.7337.96131.19140.5358.0762.30
CRA-PCN33.0425.1831.1117.6695.7265.9154.9470.0431.6133.81125.50137.1352.8059.57
GeoFormer28.0933.8131.4213.3859.7857.3350.4770.0434.9928.17131.52134.9951.9555.84
GenPC40.1236.55112.4543.9298.69149.7069.7681.4594.4068.26129.37142.4558.0086.55
SIMECO37.9034.6424.3015.88113.4972.1866.4775.3020.5041.61104.42135.8147.3560.76
PCDreamer37.7636.9645.6216.8194.2478.5291.4572.1344.8531.46135.39143.7855.4568.03
ScaleBlind (Ours)24.1225.6326.1114.7254.1446.4954.8949.0024.0422.2778.1682.6342.4641.90

ScaleBlind achieves the lowest average CD of 41.90. These results demonstrate that ScaleBlind maintains competitive completion accuracy on real-world scanned objects under the GT-scale-free setting, further supporting its generalization beyond synthetic benchmarks.

W4-1: The authors mention in the experimental details that training uses 4 GPUs, which also raises concerns about training cost.

A: We report the parameter counts and training times of representative completion methods to assess their model complexity and training costs. For consistency, all methods are trained for 300 epochs on 4× RTX 4090 GPUs, with other training settings following their respective original implementations. The parameter counts and training costs are also reported in Appendix E.8 (Table E6) of the revised paper.

MethodParams (M)Training Time (h)
AdaPoinTr32.49518.66
SVDFormer58.092471.33
CRA-PCN7.47139.92
GeoFormer58.23433.75
PCDreamer88.15135.09
SIMECO56.95772.74
ScaleBlind (Ours)64.02754.84

ScaleBlind contains 64.027M parameters and requires 54.84 hours of training on 4× RTX 4090 GPUs. Although its training time exceeds that of some baselines, it remains lower than those of SVDFormer and SIMECO, indicating a reasonable computational cost relative to existing approaches.

W4-2: The authors should provide comparisons of parameters and runtime.

A: We measure the inference time of each stage in our framework, including the parallel generation of 13 views using Nano Banana. The measured VLM generation time is 9.20 s. The detailed runtime breakdown is provided in Appendix E.8 (Table E7) of the revised paper and reported below.

StageTime (ms)
Partial point cloud rendering138.3
VLM image generation9200.0
Mask generation9.5
Voxel voting12.7
4-channel construction and FPS582.3
Loading 13 generated images20.9
Completion network forward pass21.5
Total9985.2

VLM image generation accounts for approximately 92.14% of the total inference time, indicating that the current efficiency of ScaleBlind is closely tied to VLM generation speed. We expect future advances in VLMs, particularly in generation speed, visual consistency, and reconstruction quality, to further improve both the efficiency and completion performance of our framework.

I thank the authors for their response and the additional experimental evidence. These clarifications and supplementary results have addressed some of my concerns. Therefore, I will raise my rating to 3.

We are grateful for your reconsideration and recognition of our work. Your constructive suggestions have encouraged us to extend our evaluation to the real-world Redwood dataset, provide a more comprehensive analysis of computational costs. These additions have helped us substantially strengthen the paper. Thank you again for your careful review and valuable feedback.

Comments on the paper

1 comment

We thank all reviewers for their constructive and detailed feedback. We have addressed concerns and questions through clarification and additional experiments, revised the paper by incorporating all the new results, analyses, and revisions, and updated the PDF for your reference. All the edits are marked in blue. We greatly appreciate the reviewers’ time and insightful comments, which have helped us substantially improve our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.