acceptodds
Under review as a conference paper at ICLR 2027

VIPER: Visual Prior Retrieval for Multi-Modal Point Cloud Completion

Abstract

We introduce VIPER, a simple and effective visual prior retrieval framework for multi-modal point cloud completion. Explicit image-derived 3D priors can provide structural guidance, but inaccuracies in these priors may introduce misleading geometric cues, and generating them incurs substantial computational overhead. VIPER instead treats image representations extracted by a frozen vision foundation model as queryable visual priors, using partial geometry to retrieve evidence relevant to missing or ambiguous structures. The retrieved priors are adaptively incorporated into a persistent shape representation, enabling them to guide both global shape recovery and progressive geometric refinement. On ShapeNet-ViPC, VIPER achieves state-of-the-art performance with 16.4% lower average Chamfer Distance and a 125 end-to-end inference speedup over PGNet. We further introduce OmniObject3D-Completion, a joint-category benchmark with broad object coverage for developing and evaluating multi-modal point cloud completion methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-modal Point Cloud Completion, Vision Foundation Model, Cross-modal Learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_2bxrOpenReview β†—
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents VIPER, a visual prior retrieval framework for multimodal point cloud completion. Instead of explicitly generating 3D priors, the method retrieves visual priors using pretrained visual representations and integrates them into point cloud features through a reliability gate. The authors evaluate VIPER on ShapeNet-ViPC, OmniObject3D-Completion, and KITTI, comparing its completion accuracy and inference efficiency with existing methods. The paper also introduces OmniObject3D-Completion, a new benchmark covering a broader range of object categories.

Critical strengths
  1. The paper proposes a simple and effective approach to multimodal point cloud completion. Using retrieved visual priors instead of explicitly generated 3D priors is a well-motivated design choice that improves efficiency while maintaining strong completion performance.

  2. The experiments cover multiple benchmarks, including ShapeNet-ViPC, OmniObject3D-Completion, and KITTI, providing a solid evaluation across different settings. The newly introduced OmniObject3D-Completion benchmark also covers a wider range of object categories and could be useful for future research.

  3. The paper is well written and logically organized. The motivation is clear, and the method follows naturally from the limitations of explicit 3D prior generation. The main components are also explained clearly, making the approach easy to follow.

Critical weaknesses
  1. The motivation for using qualitative examples and occlusion-impact maps in Figure 5 is not sufficiently clear. The authors should better explain how this evaluation supports their main claims.

  2. The paper lacks a discussion of the potential limitations of visual prior retrieval, particularly the scenarios where it may be less effective.

Questions for authors
  1. What motivated the evaluation approach in Figure 5, particularly the use of occlusion-impact maps to compare visual prior retrieval with explicit 3D prior generation?
  2. The proposed reliability gate controls the contribution of retrieved visual priors during feature fusion. However, it is unclear whether the learned gating values actually reflect the reliability of visual evidence or simply act as feature-wise fusion weights. Could the authors clarify this distinction, given that no explicit reliability supervision is used?
3 comments
AuthorsOpenReview β†—

Response to Reviewer 2bxr: Visual Evidence, Adaptive Gate, and Future Directions

We thank the reviewer for the constructive questions about the visual evidence analysis and the gate's interpretation.

Response to W1 and Q1 (why Figure 5 uses occlusion-impact maps). Figure 5 compares PGNet's explicit image-derived 3D prior and final completion with VIPER's completion and an input-sensitivity map. For the map, we hold the partial point cloud fixed, mask local image regions in turn, and measure how much the completion error increases. Regions with larger increases have more influence on the prediction under this perturbation. This makes the visual input's contribution to VIPER's recovered geometry inspectable even though VIPER does not generate an intermediate 3D prior. In the illustrated cases, errors in PGNet's intermediate prior coincide with completion artifacts, while VIPER recovers more faithful geometry and is sensitive to image regions associated with the missing structures. This sensitivity is consistent with VIPER's design: a frozen DINOv3 encoder supplies visual tokens, and point-conditioned top-K retrieval selects priors relevant to the partial geometry. These qualitative observations, together with the corresponding completion results, support the rationale for incorporating retrieved visual priors into point-cloud completion; stage-wise comparisons provide complementary quantitative evidence.

Revised manuscript: πŸ“ Section 3.4, page 8, lines 400–412; πŸ“ Appendix D.1, Table 10, page 21, lines 1122–1127; and πŸ“ Figure 8, page 22, lines 1161–1187.

Response to W2 (scope and future directions). Our current evaluation focuses on object-level completion from paired RGB images and partial point clouds, with quantitative results on synthetic benchmarks and qualitative transfer to KITTI LiDAR. A natural next step is to study point-conditioned visual-prior retrieval in cluttered, dynamic, and large-scale outdoor scenes, where occlusion and variable camera–LiDAR alignment make visual evidence more challenging to use. This direction is particularly relevant to multi-sensor perception for autonomous driving and mobile robotics: camera and LiDAR observations across viewpoints and time could jointly support reconstruction of surrounding geometry for perception and navigation. Beyond object-level completion, multi-view or temporal retrieval of visual foundation-model priors could support scene reconstruction and multi-modal 3D generation. We have reorganized Appendix F to discuss this broader scope and these future directions.

Revised manuscript: πŸ“ Section 3.3, Figure 4, pages 7–8, lines 373–381; and πŸ“ Appendix F, page 26, lines 1350–1360.

Response to Q2 (what the gate represents). The gate is a learned channel-wise fusion weight, conditioned on each point token and its retrieved visual message. It is trained through the completion objective without explicit reliability supervision, so its values should be understood as adaptive fusion weights rather than calibrated reliability scores. Its role is to adjust the strength of visual-prior injection for different point tokens and feature channels. We compare full VIPER with the variant without the learned injection gate in two settings:

Dataset / settingVariantGateCD ↓F-Score ↑
ShapeNet-ViPC (Cabinet; single category)VIPERβœ“0.9550.881
ShapeNet-ViPC (Cabinet; single category)w/o Adaptive Injection Gateβ€”1.0240.868
OmniObject3D-Completion (85-category joint training)VIPERβœ“3.0540.470
OmniObject3D-Completion (85-category joint training)w/o Adaptive Injection Gateβ€”3.4290.423

The dash denotes removal of the learned injection gate; in OmniObject3D-Completion variant D, the projected visual prior is added directly to the point tokens. Cabinet uses CD in units of 10^-3 and [email protected]; OmniObject3D-Completion uses CD in units of 10^-4 and [email protected]. The OmniObject3D-Completion results are sample averages under the view-based split. Full VIPER performs better in both the single-category test and the broader joint-category benchmark, which extends evaluation from ShapeNet-ViPC's limited category coverage to 85 categories. These results support the effectiveness of adaptive gate design for visual-prior integration across both settings.

Revised manuscript: πŸ“ Section 2.2, page 4, lines 168–183; πŸ“ Section 3.5, page 9, lines 448–466; and πŸ“ Appendix E.1, page 23, lines 1211–1236.

I thank the authors for their response, which has addressed my concerns. After reading the other two reviewers' comments and the authors' replies, I have decided to raise my score.

AuthorsOpenReview β†—

We Are Truly Grateful for Your Thoughtful Review

Dear Reviewer 2bxr,

Thank you sincerely for taking the time to read our response and for engaging so carefully with our work. Your questions helped us clarify what the visual evidence analysis shows, explain the role of the adaptive gate more precisely, and develop a fuller discussion of future directions.

We are grateful for the care and thought you have brought to this review. Your feedback has made the revised manuscript clearer and stronger, and we truly appreciate the opportunity to address your concerns.

Sincerely,
All Authors of Submission 816

Reviewer_QUzZOpenReview β†—
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a multimodal point cloud completion framework that extracts image features using a pre-trained vision model and integrates them with point cloud features via gated matching, thereby leveraging the complementary nature of image and point cloud information. Compared to existing image-to-3D approaches, this method significantly reduces inference costs. Furthermore, it demonstrates that when image-to-3D priors are unreliable, there is no need to explicitly convert images into 3D priors. Experimental results show that the proposed method outperforms existing state-of-the-art techniques. The effectiveness of the approach was further validated through training and testing on the OmniObject3D-Completion dataset.

Critical strengths
  1. The motivation is clear, and the proposed use of visual priors is well justified. The overall architecture is coherent, allowing visual information to contribute throughout the coarse-to-fine completion process.
  2. The experimental results demonstrate the effectiveness of the proposed method, with consistent performance improvements.
  3. Inference efficiency is another important strength of the method. For example, it achieves a 125Γ— inference speedup while using approximately one twenty-fourth as much GPU memory as PGNet.
  4. The paper also introduces a multimodal dataset, providing an additional contribution to the research community.
Critical weaknesses

1.The ablation studies are limited to a single object category. It is not clearly stated whether the removal of a specific design component affects the implementation of others, nor is it explained what mechanism is used as a replacement when a component is removed. Moreover, the performance gaps between the various ablation configurations are marginal, making it difficult to discern the specific contribution of each module.

2.The design of the ablation studies must be improved: the paper should provide a module-by-module analysis, explore combinations of different design components, and explicitly state the replacement mechanisms used when specific components are omitted.

3.It is recommended to evaluate the performance when DINOv3 is replaced by a standard encoder to verify the necessity of using DINOv3.

4.The innovation of the proposed method is limited, as the core design elements are already established techniques widely used in the field of point cloud completion. The integration of visual information remains entirely implicit: retrieved visual evidence is used solely to enhance the feature representations of point tokens, while subsequent global shape generation and the two-stage refinement process occur exclusively within the enhanced point-token space. Although the selection of visual tokens implies that these features are complementary to the point cloud, the implementation lacks an explicit representation of this complementarity. Consequently, the mechanism does not differ significantly from existing multimodal point cloud completion methods.

Questions for authors

See weakness.

5 comments
AuthorsOpenReview β†—

Response to Reviewer QUzZ: Visual Prior Retrieval and Cross-Category Ablations

We thank the reviewer for the detailed suggestions. We have expanded the ablations beyond Cabinet and clarified what each variant changes.

Response to W1–W3 (cross-category ablations, replacements, and image encoder). The new experiment covers all 85 OmniObject3D-Completion categories under the view-based split. Full VIPER uses frozen DINOv3, relevance-based Top-K retrieval (K=32), the adaptive injection gate, and decoder token guidance. A replaces DINOv3 with a ResNet-50 trained from scratch. B selects 32 visual tokens at random. C attends to all visual tokens. D removes the learned injection gate and directly adds the projected visual message to the point tokens. E retains the prior-enhanced coarse seed and global feature but does not reuse the enhanced tokens during refinement. F combines D and E. Except for A, the image encoder remains DINOv3; D–F also retain relevance-based Top-K retrieval. CD is in units of 10^-4 and F-Score is evaluated at 0.01; each entry lists the sample average / category average.

VariantImage encoderSelectionGateDecoder guidanceCD ↓ (sample / category)F-Score ↑ (sample / category)
VIPERDINOv3Top-Kβœ“βœ“3.054 / 3.0360.470 / 0.479
AResNet-50, trained from scratchTop-Kβœ“βœ“4.392 / 4.3570.359 / 0.368
BDINOv3Random-Kβœ“βœ“3.512 / 3.4960.409 / 0.417
CDINOv3Denseβœ“βœ“3.552 / 3.5310.408 / 0.416
DDINOv3Top-Kβ€”βœ“3.429 / 3.4090.423 / 0.431
EDINOv3Top-Kβœ“β€”3.316 / 3.2950.436 / 0.444
FDINOv3Top-Kβ€”β€”4.132 / 4.1000.363 / 0.372

In the Decoder guidance column, β€” means that prior-enhanced point tokens are not reused during refinement; they still condition the coarse seed and global feature. The effects are consistent on both averages. Relative to full VIPER, the sample-averaged CD increases by 0.458 with random selection (B), 0.498 with dense attention (C), 0.375 without the learned injection gate (D), and 0.262 without decoder token reuse (E). Combining the last two changes (F) increases CD by 1.078 and lowers F-Score from 0.470 to 0.363. These results distinguish the contributions of selection, adaptive injection, and token reuse beyond a single category. Replacing DINOv3 with the scratch-trained ResNet-50 (A) increases CD from 3.054 to 4.392 and lowers F-Score from 0.470 to 0.359. This substantial degradation shows the importance of DINOv3-derived visual priors to VIPER's point-conditioned retrieval in the tested configuration: retaining Top-K selection, the gate, and decoder guidance with a scratch-trained ResNet-50 does not recover the same completion accuracy. See the revised manuscript: pages 23–24, Appendix E.1, lines 1209–1243 (variant definitions, Table 11, and analysis); and page 9, Section 3.5, lines 435–466 (Cabinet ablation).

Response to W4 (novelty and visual–geometric complementarity). We agree that attention and coarse-to-fine decoding are established components. VIPER's distinction lies in how it obtains and uses foundation-model priors. A 3D foundation-model-based approach such as PGNet first turns the image into an explicit 3D hypothesis and then corrects that hypothesis using the partial point cloud. Errors in the intermediate geometry can affect the correction stage. VIPER instead keeps the image evidence in the feature space of a frozen vision foundation model and retrieves it according to the observed geometry, without generating an intermediate 3D shape.

The relationship between the two modalities is more specific than uniform feature fusion. An image can contain background and regions that are uninformative for a particular local structure. For each point query, VIPER computes relevance scores over visual tokens in a learned shared space, selects a point-specific Top-K set, and adaptively controls the injection of the retrieved message. The relevance scores, selected tokens, and gated updates operationalize point-conditioned visual–geometric complementarity, without requiring a separately supervised complementarity target. The resulting prior-enhanced point tokens condition coarse shape generation and are reused during both refinement stages.

The controlled ablations support this distinction: on OmniObject3D-Completion, relevance-based Top-K achieves a sample-averaged CD of 3.054, compared with 3.512 for Random-K and 3.552 for dense attention, while retaining the same downstream components. The matched dense comparison also shows lower inference latency for Top-K (25.0 versus 26.4 ms per sample). These results support the contribution of selective visual-prior retrieval beyond a generic cross-modal attention block. See the revised manuscript: pages 23–24, Appendix E.1, lines 1209–1243; and pages 24–25, Appendix E.4, lines 1283–1304.

Thank you for your response. Although the ablation studies have been improved, the discussion of the paper's novelty remains largely unchanged from the original submission. Consistent with the concerns raised by other reviewers, I remain unconvinced that the proposed top-k approach, without additional guidance, can substantially outperform dense attention mechanisms. The top-k strategy itself is relatively common in this field, and its advantages in terms of performance and computational efficiency are not sufficiently established. In particular, the additional sorting operation may introduce computational overhead and potentially reduce inference speed. I therefore encourage the authors to further clarify how their implementation of the top-k mechanism differs from existing approaches and what specific advantages these differences provide.

Thank you for noting the improvements in our ablation studies and for this constructive follow-up. Your comments have encouraged us to explain more clearly why point-conditioned visual retrieval helps completion and what it costs in practice.

Intuition and mechanism. Not all information in an input image is useful for completing a particular partial point cloud. The image contains background, and even object regions may differ in relevance to individual point queries. Dense attention allows every query to aggregate information from all visual tokens, so many weakly relevant tokens can collectively dilute the evidence most useful to that query. In VIPER, the partial geometry provides the guidance: each point token queries a learned shared point–image space, ranks frozen DINOv3 visual tokens by cosine similarity, and aggregates its 32 selected tokens. The retrieved message is then adaptively injected, and the enhanced point tokens are reused during refinement. This is how visual-prior selection participates in the full completion process.

Controlled quantitative evidence. Following Reviewer 6Vxf’s request for a controlled Top-K/Dense comparison, we matched the frozen encoder, projections, gate, decoder, and training budget, changing the visual-token selection rule. Under these conditions, Top-K reduces CD relative to Dense by 13.9% on Cabinet and 14.0% on OmniObject3D-Completion; F-Score rises from 0.851 to 0.881 and from 0.408 to 0.470, respectively. Random-K also underperforms relevance-based Top-K at the same retrieval size. Together, these results support selecting evidence according to the point query, beyond merely limiting the token count.

New qualitative evidence. We have now added a comparison on OmniObject3D-Completion to examine what accompanies this accuracy difference. In the paired image-pixel response maps in Figure 12, Top-K concentrates responses more strongly on object regions, whereas Dense also responds broadly to the surrounding background. In several examples, including Toy Bus and Eraser, the broader Dense responses accompany less faithful completed geometry. These observations provide a visual account consistent with the controlled quantitative results.

Measured inference cost. Ranking and selecting Top-K tokens incurs overhead. We therefore assess efficiency using measured inference latency for the complete variants, rather than assuming that sparse attention must be faster. Top-K takes 30.7 versus 31.1 ms per sample on Cabinet and 25.0 versus 26.4 ms on OmniObject3D-Completion; peak GPU memory is 0.42 versus 0.46 GiB in both settings. These latency differences are modest. The accuracy improvement does not come with an increase in measured inference cost relative to the matched Dense variant. The much larger speedup over PGNet arises from the overall framework avoiding explicit image-to-3D prior generation and should not be attributed to Top-K alone.

Benchmark contribution. We also appreciate your earlier recognition of OmniObject3D-Completion as a contribution. It provides joint training across 85 categories and object-disjoint evaluation of the full method. The controlled Top-K/Dense comparison uses its view-based setting; the benchmark’s broader evaluation is a separate contribution alongside the retrieval design.

Revised manuscript: πŸ“ Appendix E.1, Table 11, page 23; πŸ“ Appendix E.4, Top-K/Dense comparison, pages 24–25, including Table 12 on page 25; and πŸ“ Appendix G, Figure 12, page 26, lines 1392–1403.

Reviewer_6VxfOpenReview β†—
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a multimodal point cloud completion method, VIPER. It completes partial point clouds using frozen DINOv3 image features through point-conditioned top-K attention and progressive refinement. Compared with previous method PGNet, the paper reports a 16.4% reduction in average CD on ShapeNet-ViPC and approximately 125Γ— faster end-to-end inference. It also introduces the 85-category OmniObject3D-Completion benchmark. The consistent improvements on the established benchmark and the favorable balance between accuracy and efficiency are the main strengths of the submission.

Critical strengths
  1. The paper writing is easy to follow with sufficient details and clear figures.
  2. The idea of retrieving top-k visual tokens for cross-modal attention sounds novel.
Critical weaknesses
  1. The technical novelty is limited. The overall framework consists of two stages, including cross-modal attention and coarse-to-fine decoding, which are widely used for point cloud completion tasks. The idea of retrieving top-k visual tokens for each point to compute cross-attention sounds novel but is incremental.

  2. The independent benefit of the retrieval mechanism is not sufficiently established. The top-K ablation in Section 3.5 and Appendix E.2 replaces relevance-based selection with random selection of 32 tokens. This demonstrates the value of relevance-based selection over random selection, but not its advantage over dense cross-attention. A direct comparison retaining the same DINOv3 encoder, projections, gates, decoder, and training budget is missing, leaving the specific contribution of selective retrieval insufficiently isolated.

  3. No results on unseen objects are reported to demonstrate the generalization capability. In the main text, the inference results on KITTI cars are shown based on the completion models trained on the car category of the 3D-EPN dataset, while the previous methods (such as XMFnet, EGIInet, and GAPM) can complete point cloud on unseen categories. Appendix B explicitly uses different views of the same objects for training and testing. Consequently, the complete geometry of a test object is available during training through its other views. This protocol evaluates completion from new views of known objects, but does not establish unseen-instance generalization. Since the benchmark is a major claimed contribution, an object-disjoint evaluation is important for establishing its broader value.

  4. Insufficient efficiency comparison with previous methods. The authors only report the inference cost comparison between PGNet and VIPER, but they did not compare the proposed VIPER with other point cloud completion methods (such as XMFnet, EGIInet, and GAPM).

Questions for authors
  1. How does top-K attention compare with dense cross-attention in completion accuracy and inference latency when all other components and the training budget are held fixed? Does the conclusion hold beyond Cabinet?
  2. Do VIPER’s improvements over the main baselines on OmniObject3D-Completion persist under an object-disjoint train/test split?
7 comments
AuthorsOpenReview β†—

Response to Reviewer 6Vxf (1/2): Method and Retrieval Mechanism

Response to W1: We thank the reviewer for this comment. We agree that cross-modal attention and coarse-to-fine decoding are established components. VIPER's contribution lies in how it obtains and uses foundation-model priors. PGNet, a 3D foundation-model-based method, first generates an explicit image-derived 3D hypothesis and then corrects it with the partial point cloud; inaccuracies in that intermediate geometry can affect completion, and generating it incurs substantial cost. VIPER instead keeps visual evidence in the feature space of frozen DINOv3. Each point query selects relevant visual tokens through Top-K retrieval, a learned gate controls their injection, and the prior-enhanced point tokens guide both coarse shape generation and subsequent refinement. This point-conditioned use of visual priors distinguishes VIPER from simply applying generic cross-modal attention. On ShapeNet-ViPC, VIPER achieves 16.4% lower average CD than PGNet, with 125Γ— faster end-to-end inference and approximately 24Γ— less GPU memory. The added controlled retrieval and cross-category ablations further examine these design choices. Further details are provided in the revised manuscript: pages 8–9, Section 3.4, Table 3 (lines 415–433); pages 23–24, Appendix E.1, Table 11 (lines 1209–1243); and pages 24–25, Appendix E.4, Table 12 (lines 1283–1304).

Response to W2 and Q1 (top-K versus dense attention). We added a controlled comparison that changes only the visual-token selection rule: both variants use the same frozen DINOv3 encoder, projections, injection gate, decoder, and training budget. Dense attention uses all visual tokens; top-K selects the 32 most relevant tokens for each point query. The OmniObject3D-Completion results use the 85-category view-based split. CD is reported in units of 10^-3 for Cabinet and 10^-4 for OmniObject3D-Completion; F-Score thresholds are 0.001 and 0.01, respectively.

Dataset / categoryVariantCD ↓F-Score ↑Trainable Params. (M)Frozen Params. (M)Inference Latency (ms) ↓GPU Memory (GiB) ↓
ShapeNet-ViPC (Cabinet)VIPER (Top-K)0.9550.88153.7721.6130.70.42
ShapeNet-ViPC (Cabinet)VIPER (Dense)1.1090.85153.7721.6131.10.46
OmniObject3D-CompletionVIPER (Top-K)3.0540.47053.7721.6125.00.42
OmniObject3D-CompletionVIPER (Dense)3.5520.40853.7721.6126.40.46

Top-K reduces CD by 13.9% on Cabinet and 14.0% on OmniObject3D-Completion, improves F-Score by 0.030 and 0.062, respectively, and has lower measured latency and GPU memory. This directly tests selective retrieval beyond Cabinet. Further details are provided in the revised manuscript: pages 24–25, Appendix E.4, analysis on page 24 (lines 1289–1295) and Table 12 on page 25 (lines 1297–1304); and page 23, Appendix E.1, Table 11.

AuthorsOpenReview β†—

Response to Reviewer 6Vxf (2/2): Generalization and Efficiency

We thank the reviewer for these helpful comments on unseen-object generalization and efficiency.

Response to W3 and Q2 (unseen-object generalization). We agree that the original view-based OmniObject3D-Completion split allows different views of one object in training and testing. We now evaluate a separate object-disjoint split, in which no test object's complete geometry occurs in training. This tests new instances within the benchmark's category vocabulary; it is distinct from testing entirely unseen categories. The new object-disjoint aggregate results are reported below; VIPER achieves the best CD and F-Score among the compared methods. CD is in units of 10^-4 and F-Score is evaluated at 0.01.

Object-disjoint split (85-category joint training; test objects excluded from training):

MethodSample-averaged CD ↓Category-averaged CD ↓Sample-averaged F-Score ↑Category-averaged F-Score ↑
PGNet12.552511.85160.26460.2731
IAET10.099110.03230.25150.2607
XMFnet9.79279.71860.33280.3393
EGIInet9.49509.52850.33830.3449
VIPER8.34968.31880.37210.3802

Against the strongest baseline on these aggregate measures, EGIInet, VIPER reduces sample-averaged CD by 12.1% and raises F-Score by 0.0338. The category averages show the same ordering. We also added a separate test on four unseen ShapeNet-ViPC categories (Bench, Monitor, Speaker, and Cellphone): VIPER achieves 1.262 average CD and 0.843 average F-Score, versus GAPM's 1.684 and 0.802, and leads on each of the four categories. Further details are provided in the revised manuscript: page 5, Section 3.1, lines 258–264 (split definitions); page 16, Appendix C.1, lines 852–858 (object-disjoint analysis); pages 19–20, Appendix C.1, Tables 7–8 (complete per-category object-disjoint results); and page 20, Appendix C.2, Table 9 (unseen categories).

Response to W4: Thank you for suggesting a broader efficiency comparison. We additionally include XMFnet, EGIInet, and GAPM in the comparison. XMFnet and EGIInet were measured in the same experimental environment as VIPER on an NVIDIA H100 PCIe GPU. The accuracy results are eight-category averages on ShapeNet-ViPC.

Method familyMethodAvg. CD ↓Avg. F-Score ↑Trainable Params. (M)Frozen Params. (M)Inference Latency (ms) ↓GPU Memory (GiB) ↓
Traditional methodsXMFnet1.4430.79610.04014.290.139
Traditional methodsEGIInet1.2110.8369.47013.100.103
Traditional methodsGAPM†1.1070.860β€”β€”β€”β€”
Foundation-based methodsPGNet0.9260.89561.251200384710.13
Foundation-based methodsVIPER0.7740.92253.7721.6130.70.42

VIPER has a higher inference cost than XMFnet and EGIInet, but delivers substantially better completion accuracy. Relative to these two methods, VIPER reduces average CD by 46.4% and 36.1%, respectively, while increasing F-Score by 0.126 and 0.086. Compared with GAPM, VIPER reduces CD by 30.1% and increases F-Score by 0.062.

† GAPM's accuracy is taken from its published results. As GAPM is closed-source and its paper reports no inference cost, we cannot benchmark it under the same conditions; its efficiency entries are left blank.

Further details are provided in the revised manuscript: pages 8–9, Section 3.4, Table 3 and lines 415–433.

I didn't expect that top-k's performance wouldn't be as good as the result of dense attention. I think dense attention should capture more information and will lead to better performance. But why is the CD even higher? Can the authors provide a more detailed analysis of the attention and more explanation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.