acceptodds
Desk rejected at ICLR 2027

GEAR: Versatile Transductive Diffusion for Training-Free 3D Object Retrieval

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

Abstract

Deep 3D embeddings derived from pretrained models have demonstrated promising retrieval performance. However, achieving optimal results typically requires downstream adaptation with labeled 3D training data, which incurs significant annotation costs and limits their applicability in evolving real-world scenarios. In this paper, we introduce GEAR, a training-free test-time adaptation framework that efficiently performs transductive diffusion to improve the embedding discriminativeness without requiring labeled 3D data for adaptation. Specifically, GEAR only exploits the proximity information among unlabeled gallery objects and diffuses in a transductive manner. In this way, we enforce a cohesive embedding space for gallery samples suitable for retrieval. Unlike conventional transductive approaches that require a batch of queries for joint adaptation, GEAR enables single-query adaptation during inference by leveraging the refined gallery representations, making it applicable to standard online retrieval scenarios. Furthermore, we introduce an iterative anchored diffusion strategy to improve retrieval ranking while preserving the original query evidence. Extensive experiments on 3D object retrieval benchmarks demonstrate that GEAR effectively boosts the performance of various pretrained representations. Additional evaluations on sketch-based and scene-level retrieval further verify its flexibility and generalizability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D object retrieval, test-time refinement, training-free adaptation, feature refinement, graph diffusion, representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_L38MOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GEAR, a training-free framework for 3D object retrieval that exploits unlabeled gallery structure through transductive diffusion. It adopts an offline-online pipeline consisting of gallery centering, multi-scale neighborhood refinement, single-query calibration, and anchored score diffusion. GEAR enables independent query processing without downstream training or gradient optimization. Experiments on four open-set 3D retrieval benchmarks demonstrate consistent improvements across multiple pretrained backbones, with additional evaluations on point-cloud, cross-modal, and scene-level retrieval tasks.

Critical strengths
  1. GEAR addresses a practical setting where gallery objects are available in advance while queries arrive independently. Its offline-online design avoids batch-query dependencies and enables efficient single-query retrieval without downstream training.
  2. GEAR consistently improves frozen representations across CLIP, DINOv2, and DINOv3. It also achieves competitive or superior performance compared with supervised adaptation methods such as DEC (CVPR 2026) and transductive methods such as TeDA in the reported settings.
  3. The paper provides extensive ablation studies, parameter sensitivity analyses, statistical evaluations, and efficiency measurements. The unified hyperparameter configuration across datasets and backbones is appreciated. Additional experiments on different retrieval tasks further demonstrate the flexibility of the framework.
Critical weaknesses
  1. The main components of GEAR are closely related to established techniques, including query expansion, database-side augmentation, and graph diffusion. However, comparisons with these classical training-free retrieval methods are missing. The related work also overlooks several recent studies, such as Source-Free Model Adaptation for Unsupervised 3D Object Retrieval (TVCG 2025), Uni-Adapter (AAAI 2026), and PASR (CVPRF 2026), although their task settings differ from GEAR.
  2. The proposed anchored diffusion provides only marginal improvements (+0.34 mAP points), while ordinary diffusion achieves better mAP (68.74 vs. 68.53) at the default L=2 (Table 9). Similarly, Multi-K refinement slightly underperforms single-K refinement (66.66 vs. 66.71 mAP, Table 7), despite additional computation. These results raise concerns about the necessity of these components.
Questions for authors
  1. Could the authors provide comparisons against established training-free retrieval enhancement methods, such as QE, DBA, and conventional graph diffusion or re-ranking, using identical frozen descriptors? What are the main methodological advantages of GEAR over existing offline-online diffusion frameworks?
  2. Given that ordinary diffusion outperforms iterative anchoring at L=2, what justifies choosing anchored diffusion as the default strategy? Similarly, since Multi-K refinement does not consistently outperform single-K refinement, could the authors provide a more detailed accuracy-efficiency analysis to justify its additional computational cost?
  3. Could the authors clarify whether GEAR and DEC follow identical rendering, feature extraction, and evaluation protocols? Can additional backbone-matched comparisons be provided, particularly with stronger DINO representations? Furthermore, would applying GEAR to DEC-adapted representations yield additional improvements, thereby supporting its claimed plug-and-play generality?
8 comments
AuthorsOpenReview ↗

Response to Reviewer L38M [1/4]

Weakness 1 and Question 1: Classical Retrieval Baselines and Related Work

Thank you for your thorough and constructive review, which helps us clarify and strengthen our work.

1. Comparisons with Classical Retrieval Methods. We compare GEAR with query expansion, database-side augmentation, standard graph diffusion, and a classical full pipeline combining gallery centering, weighted AQE/DBA, and restart-based graph propagation. All methods use identical frozen descriptors, dataset splits, and evaluation protocols. Classical hyperparameters follow strict leave-one-dataset-out validation, while GEAR retains its submitted fixed configuration. The table reports macro mAP.

MethodCLIPDINOv2DINOv3
Frozen descriptors56.3860.9860.64
Gallery centering only60.2163.8363.15
Weighted AQE + DBA63.7566.5965.99
Standard global diffusion65.3666.9966.05
Classical full pipeline63.9966.2765.20
GEAR (Ours)65.6568.5367.96

Classical refinements substantially improve frozen descriptors, confirming the value of gallery structure. GEAR achieves the highest macro mAP among the reported baselines, outperforming standard global diffusion by +0.29, +1.54, and +1.91 points, respectively. Against the classical full pipeline, gains reach +1.66, +2.26, and +2.76 points, with all paired 95% confidence intervals above zero. Individual baseline variants and protocols are detailed in Appendix D.6.

2. Differences from Existing Offline-Online Diffusion Frameworks. Existing frameworks, such as the 2019 AAAI study by Yang and colleagues, mainly exploit database-side graph precomputation for efficient online ranking. GEAR instead uses offline computation to refine gallery representations and adapts each incoming query against the refined gallery. This distinction involves two mechanisms.

First, gallery refinement iteratively recomputes neighborhoods, controls residual updates through local agreement, and fuses independently refined multi-scale representations. The resulting embeddings and transition graph are cached offline as a reusable structural prior.

Second, rather than directly propagating frozen query-gallery similarity scores, GEAR recalibrates each query against the refined gallery through an agreement-dependent residual update before anchored score propagation. This supports independent, query-specific adaptation without downstream optimization or interactions between queries.

To isolate this mechanism, we conducted a controlled module-replacement experiment under a shared diffusion-free setup. Holding gallery representations fixed, replacing conventional query refinement with GEAR's recalibration improves mAP by 0.44–0.55 points, with positive paired 95% confidence intervals. This directly supports the value of gallery-conditioned query recalibration.

3. Addition of Recent Related Work. We have incorporated all three suggested studies. Source-Free Model Adaptation for Unsupervised 3D Object Retrieval (TVCG 2025) studies prototype-based source-free adaptation and distillation; PASR (CVPR Findings 2026) addresses single-view image-to-3D retrieval using test-time pose optimization; and Uni-Adapter (AAAI 2026) explores training-free online adaptation of 3D vision-language models through dynamic class-prototype caches and graph-based label smoothing. The first two are discussed in Section 2.1 and Uni-Adapter in Section 2.2, with their task settings distinguished from GEAR.

Overall, these additional comparisons substantiate GEAR's performance under matched frozen representations, clarify its gallery-refinement and single-query adaptation mechanisms, and address the missing recent literature.

AuthorsOpenReview ↗

Response to Reviewer L38M [2/4]

Weakness 2 and Question 2: Justification of Multi-K and Anchored Diffusion [1/2]

Regarding Multi-K: Neighborhood-Scale Heterogeneity and Accuracy-Efficiency Trade-off

Thank you for your constructive comments. Our choice of Multi-K is motivated by heterogeneous neighborhood structures within real-world galleries, where incoming queries may target objects with different amounts of reliable neighboring support. We examine whether multi-scale refinement offers practical benefits that justify its additional offline computation.

1. Neighborhood-scale heterogeneity. Different gallery regions favor different refinement scales. Multi-K independently refines representations at multiple scales before fusion, rather than relying on a single K=20K=20 trajectory. Density-stratified analyses reveal the following preferences:

Gallery DensityLowMediumHigh
K=5K=5311
K=10K=10650
K=20K=20025
K=50K=50346

Number of backbone-dataset settings where each Single-K achieves the highest mAP. Density groups are determined from frozen gallery features; labels are used only for post-hoc evaluation.

Smaller scales are preferred in 9/12 low-density settings, while larger scales are favored in 11/12 high-density settings. A complementary mixed-gallery experiment confirms that these opposing preferences coexist within the same gallery: K=10K=10 outperforms K=50K=50 by +0.964 mAP points in the 24-object group, whereas K=50K=50 outperforms K=10K=10 by +0.884 points in the high-capacity group. These findings show that similar aggregate mAP can conceal distinct local refinement requirements.

2. Cross-dataset scale selection and practical applicability. In heterogeneous asset libraries, future queries may originate from any gallery region, making reliance on a single neighborhood scale potentially restrictive. Multi-K integrates complementary scale-specific refinement trajectories into a unified gallery representation without requiring target relevance labels for regional scale selection. Under leave-one-dataset-out (LODO) validation, Multi-K improves mAP over the selected Single-K by +0.134, +0.149, and +0.220 points for CLIP, DINOv2, and DINOv3, with a positive paired 95% confidence interval on DINOv3 [+0.098, +0.347].

3. Accuracy-efficiency trade-off. The central consideration is whether multi-scale refinement's additional computation affects practical retrieval efficiency. We compare Single-K and Multi-K using matched implementations on ModelNet40 with 9,329 gallery objects and DINOv2.

MethodmAP (%)Offline PreparationOnline P50 (ms)Final Cache
Single-K (K=20K=20)74.4112.56 s4.9929.28 MiB
Multi-K74.4058.07 s4.9029.27 MiB

Multi-K requires approximately 4.6× the offline preparation time, but this additional computation does not increase online inference cost. The independently refined representations are fused into a single gallery representation offline. Once cached, each query uses this unified representation without executing multiple scale-specific refinement branches. Consequently, online latency remains approximately 5 ms per query, with virtually identical cache requirements.

This separation matches GEAR's practical retrieval setting: the gallery is prepared in advance, while queries arrive independently over time. For relatively stable galleries serving repeated queries, multi-scale preparation is performed once and reused across subsequent retrievals. In return, the gallery representation integrates structural information from multiple neighborhood ranges without relying on a single scale for all objects. The trade-off is less favorable for frequently changing galleries that require repeated preparation.

Overall, Multi-K deliberately allocates additional offline computation to integrating heterogeneous neighborhood structures into a reusable representation while preserving efficient online single-query retrieval. The density-stratified and cross-dataset analyses support this design, even though Multi-K does not consistently improve aggregate mAP over K=20K=20. We will incorporate these analyses and computational measurements into the revised Appendix.

AuthorsOpenReview ↗

Response to Reviewer L38M [3/4]

Weakness 2 and Question 2: Justification of Multi-K and Anchored Diffusion [2/2]

Regarding Anchored Diffusion: Top-Rank Preservation and Robustness

Thank you for your valuable feedback. Our choice of anchoring reflects a deliberate trade-off between aggregate ranking accuracy and top-rank reliability. In practical single-query 3D asset retrieval, users typically inspect the highest-ranked results first. Although graph propagation can improve overall ranking, unreliable connections may also overturn initially correct top-ranked results. Anchored diffusion addresses this by reintroducing the initial query scores at each propagation step, limiting harmful score drift while retaining the benefits of gallery structure. We conducted controlled experiments to evaluate this design choice.

1. Top-Rank Preservation. We compare ordinary and anchored diffusion at the submitted fixed L=2L=2.

BackboneInitial P@1Ordinary P@1Anchored P@1Ordinary mAPAnchored mAP
CLIP84.8478.7984.6065.8565.65
DINOv287.1482.4686.7568.7468.53
DINOv387.7182.4787.9568.0767.96

Anchoring improves P@1 over ordinary diffusion by 5.81, 4.30, and 5.48 points, with all paired 95% confidence intervals above zero, while sacrificing only approximately 0.1–0.2 mAP points. Ordinary diffusion substantially reduces the initially strong P@1, whereas anchoring largely preserves it. Per-query analysis shows that anchoring protects 56, 39, and 45 initially correct top-ranked results disrupted by ordinary diffusion. Meanwhile, anchored diffusion still improves mAP over no diffusion by 0.35, 0.35, and 0.38 points. Thus, the marginal mAP trade-off yields a substantial top-rank reliability benefit while retaining the overall gains from graph propagation.

2. Robustness to Unreliable Graph Connections. To examine robustness to unreliable graphs, we keep the refined gallery, initial scores, and L=2L=2 fixed, perturbing only the final graph. We evaluate random edge rewiring and wrong-class edge rewiring at a 25% perturbation rate using three fixed seeds.

BackboneRandom Rewiring: Ordinary / AnchoredWrong-Class Rewiring: Ordinary / Anchored
CLIP7.97 / 0.698.72 / 0.79
DINOv27.81 / 0.648.57 / 0.71
DINOv37.86 / 0.698.51 / 0.73

Values report mAP degradation relative to each method's clean graph; lower is better. All six paired confidence intervals for the reduction in degradation are positive.

Anchoring reduces mAP degradation by 7.17–7.93 points under these controlled perturbations, supporting its role in limiting the influence of erroneous graph connections on query-specific scores. Wrong-class rewiring is solely diagnostic; class labels are never required during inference.

3. Stability across Propagation Depths. We further evaluate depths of 1, 2, 5, 10, and 20 across 12 backbone-dataset settings. At the submitted fixed L=2L=2, anchored diffusion is only 0.015 mAP points below its post-hoc optimal depth on average, compared with 0.618 points for ordinary diffusion. This indicates substantially lower depth sensitivity, which is valuable for unlabeled galleries where target-label-based depth selection is impractical.

Overall, we retain anchored diffusion because its small sacrifice in aggregate mAP provides measurable benefits in top-rank accuracy, resistance to erroneous graph propagation, and fixed-depth stability. These properties address the reliability requirements of online single-query retrieval, where preserving initially correct results is important and target-specific parameter selection is unavailable. We will incorporate the supporting analyses into the revised Appendix.

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

The article presents GEAR, a training-free framework for 3D object retrieval. It has an offline and an online part. Offline, GEAR applies two steps, GALLERY CENTERING and MULTI-K NEIGHBORHOOD REFINEMENT. Online, it applies SINGLE QUERY EMBEDDING RECALIBRATION and QUERY-GALLERY SCORE CALIBRATION based on Anchored Score Diffusion. Unlike batch-transductive methods, GEAR does not need simultaneous access to multiple queries, category labels, gradient updates, or downstream model training. Across datasets and backbones it reaches SOTA-level performance, even surpassing methods that require training.

Critical strengths
  1. A training-free 3D object retrieval framework that generalizes well. Through operations such as GALLERY CENTERING and SINGLE QUERY EMBEDDING RECALIBRATION, GEAR neatly removes the effect of visual differences across the dataset during retrieval.

  2. Broad and fairly consistent empirical gains. Quantitative experiments on CLIP, DINOv2, and DINOv3 backbones over four datasets show the framework’s superiority, and ablations show that each component contributes.

Critical weaknesses
  1. Some claims in the article are not backed by experiments. The Method section states that the statistical mean of the gallery embedding matrix (\mu_t in the paper) captures shared features at the dataset level. The text offers neither citations nor visual experiments in support of this view.

  2. The experimental basis for the Multi-K design could be stronger. Results in the main text and appendix indicate that Multi-K can achieve relatively stable performance on several datasets, and that the reliability coefficient computed from local similarity can suppress the influence of unreliable neighborhood information. That mechanism, however, mainly adjusts the strength of feature updates inside each scale branch; it cannot explain how the fixed neighborhood scale set 5, 10, 20, 50 was chosen. Each scale branch is also fused with equal weight. The existing experiments mostly report macro averages over datasets and have not fully shown how each K performs on individual datasets or under different local data densities. Since the practical value of a training-free method depends a good deal on fixed hyperparameters transferring across datasets, I suggest adding single-scale versus Multi-K comparisons for each dataset, or a grouped analysis by data characteristics such as category frequency and local sample density. It would also help to explain whether the reliability coefficient can reduce the sensitivity to scale choice. On the present evidence, Multi-K can be considered a more robust default configuration.

Questions for authors
  1. Theoretical support: Could you provide theoretical arguments or experimental evidence for some of the statements in the article, such as the statistical mean of the gallery embedding matrix, and for parameter settings such as why the multi-k features are averaged?
  2. Hyperparameter-selection protocol: Could you add theoretical support for the hyperparameter choices, for example lambda_max=0.7, multi-k={5, 10, 20, 50}, and the direct averaging over multi-k?
  3. Dataset analysis: Could you analyze, in the appendix or main text, the density of objects in different categories and the related retrieval parameters? This would give better support to the hyperparameter selection protocol.
6 comments
AuthorsOpenReview ↗

Response to Reviewer MFFi [1/3]

Weakness 1 and Question 1: Gallery Centering

Regarding Gallery Centering: Evidence for a Shared Mean Direction and Retrieval Gains

Thank you for your valuable comments. Geometrically, the normalized nonzero gallery mean identifies the direction that maximizes the average projection of gallery descriptors, without presupposing any particular visual semantics. Mean centering also has precedents in representation learning; for example, the 2019 SimpleShot study demonstrates the effectiveness of mean subtraction followed by L2 normalization for visual nearest-neighbor classification. We further conducted controlled experiments to examine whether the gallery mean captures a statistical direction shared across categories and how its removal affects retrieval.

1. Cross-class consistency and held-out validation. We performed 20 random class-disjoint splits across the evaluated benchmarks, comparing their mean directions under class-balanced averaging to control for unequal category sizes. We also projected held-out categories onto the direction estimated from the other group, using 64 random unit directions per split as controls. The following table reports DINOv2 results.

DatasetClass-balanced mean cosineHeld-out projectionRandom directions
ESB0.93830.59370.0002
NTU0.94910.48190.0002
MN400.94180.5441≈0
ABO0.85280.4481−0.0002

Despite having no overlapping categories, the two groups exhibit strongly aligned mean directions. This consistency extends to CLIP (0.968–0.987) and DINOv3 (0.872–0.964). Held-out categories also maintain positive projections along the estimated direction, whereas random-direction controls remain near zero.

These results provide direct evidence of a cross-category shared statistical direction that cannot be explained solely by category-frequency imbalance. While identifying its exact visual origin, such as dataset-specific factors or encoder-wide anisotropy, requires further study, the observed directional consistency is clear.

2. Feature-space discrimination. We further examine how centering affects feature geometry using 100,000 matched object pairs per setting, evaluated before and after centering followed by ℓ2\ell_2 normalization. We report the within-class minus between-class cosine gaps on DINOv2.

DatasetBefore centeringAfter centering
ESB0.28270.4679
NTU0.38840.5167
MN400.29650.4446
ABO0.29200.4442

Centering increases the separation gap across all four datasets, with between-class similarity decreasing more than within-class similarity. Additional geometric diagnostics show consistent improvements in several settings; for example, DINOv2-ESB hubness skewness decreases from 0.777 to 0.569.

3. Retrieval effectiveness. The geometric improvements also translate into retrieval gains. In a simplified DINOv2 pipeline without query recalibration or diffusion, centering increases four-dataset macro mAP by 2.848 points, from 60.979 to 63.827. With Single-K=20K=20 gallery refinement, centering improves mAP by 1.335 points, from 65.378 to 66.713.

We additionally isolate centering within the complete GEAR pipeline, keeping all other components and configurations unchanged.

BackboneΔmAP (pp)Paired 95% CI
CLIP+0.675[−0.142, +1.511]
DINOv2+1.719[+1.127, +2.344]
DINOv3+1.536[+0.975, +2.148]

Centering improves mAP in 11 of 12 backbone-dataset settings by point estimate, with positive macro-average confidence intervals for DINOv2 and DINOv3. The CLIP confidence interval includes zero.

Together, these results support the interpretation of μt\mu_t as a cross-category shared statistical direction and demonstrate the practical effectiveness of gallery centering, without attributing this direction to an unverified visual source.

We will incorporate the geometric interpretation and the SimpleShot citation into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer MFFi [2/3]

Weakness 2 and Questions 2–3: Multi-K Design and Hyperparameter Selection

Regarding Multi-K: Density, Fusion, and Hyperparameter Analyses

Thank you for your suggestions on per-dataset comparisons, neighborhood density, scale selection, and fusion weights. The reliability coefficient controls update strength within each scale, while the choice of neighborhood scales and fusion weights requires separate examination. We therefore conducted additional analyses across three backbones and four datasets, including controlled comparisons of scale sets, fusion strategies, and reliability coefficients.

1. Per-dataset Single-K versus Multi-K comparisons. We evaluate Single-K at K=5, 10, 20, and 50 alongside the submitted Multi-K configuration on every dataset. The following table presents DINOv2 results with identical frozen descriptors and downstream settings.

Gallery refinementESBNTUMN40ABOMacro mAP
K=5K=563.23965.29273.09370.52368.037
K=10K=1063.54965.60173.77471.02968.488
K=20K=2063.48065.29774.40770.98368.542
K=50K=5063.12164.77875.09970.37268.342
Multi-K63.30665.43774.40270.98668.533

The preferred Single-K varies across datasets: K=10K=10 performs best on ESB, NTU, and ABO, whereas K=50K=50 performs best on MN40. Multi-K does not dominate every individual dataset but provides a competitive unified configuration without selecting KK using target-gallery labels. Its macro mAP difference from K=20K=20 is only −0.009 points, with a paired 95% CI of [−0.164, +0.121], while its NDCG@100 improves from 48.014 to 48.126.

We additionally compare fixed Multi-K against Single-K selected using the other three datasets under leave-one-dataset-out (LODO) validation. Multi-K achieves macro mAP gains of +0.134, +0.149, and +0.220 for CLIP, DINOv2, and DINOv3. The strongest statistical evidence is observed on DINOv3, whose paired confidence interval excludes zero. These results support the practical value of a competitive fixed multi-scale configuration rather than universal superiority over a carefully chosen Single-K.

2. Category frequency and local-density analyses. Following your suggestion, we further group retrieval queries by gallery class frequency, class-level local feature density, and query-neighborhood similarity. The following table reports the post-hoc best Single-K for the low-, medium-, and high-value groups on DINOv2, listed in that order.

DatasetClass frequencyLocal feature densityQuery-neighborhood similarity
ESB5, 10, 205, 20, 510, 5, 20
NTU10, 10, 2010, 10, 5010, 10, 50
MN4050, 50, 5050, 50, 5050, 50, 50
ABO10, 20, 5010, 10, 5010, 10, 20

These analyses reveal heterogeneous scale preferences both across and within galleries. For instance, on ABO, the preferred K increases from 10 to 50 across increasing class-frequency groups, while the low-, medium-, and high-density groups favor K=10, K=10, and K=50, respectively. The relationship is not universally monotonic: all three MN40 groups prefer K=50.

This heterogeneity motivates covering multiple neighborhood extents rather than relying on one dataset-specific scale. Class labels are used only for post-hoc grouping and evaluation, never for GEAR inference or scale fusion.

AuthorsOpenReview ↗

Response to Reviewer MFFi [3/3]

Weakness 2 and Questions 2–3: Multi-K Design and Hyperparameter Selection (Continued)

3. Why equal-weight fusion? We compare equal averaging against small-K-biased, large-K-biased, and similarity-weighted fusion without target-label-based weight selection.

Fusion strategyCLIP mAPDINOv2 mAPDINOv3 mAP
Equal weighting (GEAR)65.65468.53367.956
Small-K biased65.63968.45567.842
Large-K biased65.67468.50367.935
Similarity-weighted65.68468.54167.950

Small-K-biased fusion uses weights proportional to 1/K1/\sqrt{K}.

No tested alternative consistently improves over equal weighting across all three backbones. Similarity weighting improves DINOv2 mAP by only +0.008 points, with a paired 95% CI of [−0.024, +0.040]. Equal fusion is therefore a simple, competitive default that avoids additional scale-specific weight selection.

For unit-normalized branch descriptors, the normalized equal-weight sum, when nonzero, minimizes the total squared Euclidean distance to the branch descriptors over the unit sphere. This provides a geometric interpretation of equal fusion without implying that it is uniquely optimal for retrieval.

4. Why the scale set {5,10,20,50}\{5,10,20,50\}? Our design covers relatively local through broader neighborhoods while keeping the scale set fixed across galleries. To examine whether the submitted range and number of branches are reasonable, we compare three-, four-, and five-scale variants under the same DINOv2 retrieval protocol.

Scale setMacro mAPNDCG@100ANMRR ↓Relative offline cost
{5,10,20}\{5,10,20\}68.46548.10335.1860.26–0.55×
{5,10,20,50}\{5,10,20,50\} (GEAR)68.53348.12635.1341.00×
{5,10,20,50,100}\{5,10,20,50,100\}68.51148.09235.1911.76–2.19×
{10,20,30,40}\{10,20,30,40\}68.54048.04835.1070.99–1.13×

Offline cost is normalized to the submitted four-scale configuration; ranges are reported across the evaluated benchmarks.

Removing K=50K=50 reduces macro mAP by approximately 0.07 points, with the largest loss on MN40 (−0.425 points), although the three-scale variant substantially reduces offline cost. Extending to K=100K=100 increases offline cost by 1.76–2.19× without improving overall mAP or NDCG@100. Concentrating four scales in {10,20,30,40}\{10,20,30,40\} produces comparable mAP (+0.007 points) but lower NDCG@100 than the submitted set.

These results support the fixed scale set of 5, 10, 20, and 50 as a practical accuracy-cost compromise, rather than a uniquely optimal choice.

5. Reliability and hyperparameter sensitivity. We test whether adaptive reliability reduces sensitivity to neighborhood scale by comparing it with a fixed coefficient matched to its average update strength. On DINOv2, adaptive reliability reduces the standard deviation of macro mAP across K=5, 10, 20, and 50 from 0.270 to 0.196 points, a reduction of approximately 27%. The paired 95% interval for the adaptive-minus-fixed difference is [−0.131, −0.002]. CLIP and DINOv3 show the same directional trend, although their intervals include zero.

Thus, adaptive reliability can moderate sensitivity to K without selecting the neighborhood scale or assigning fusion weights.

Finally, we evaluate λmax⁡∈{0,0.3,0.5,0.7,0.9,1.0}\lambda_{\max}\in\{0,0.3,0.5,0.7,0.9,1.0\}. On DINOv2, the submitted value 0.7 achieves the highest observed macro NDCG@100 of 48.126, although larger values yield slightly higher mAP. This reflects a metric trade-off rather than a universal optimum.

In summary, the new experiments support the motivation for a fixed multi-scale configuration: neighborhood-scale preferences vary across gallery structures, while equal fusion and adaptive reliability offer simple, competitive choices without target-label-dependent configuration. We will incorporate the per-dataset comparisons, density-group analyses, fusion and scale-set ablations, and available statistical evidence into the revised Appendix.

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

The paper introduces a training-free framework for multi-view image-based 3D object retrieval that improves pretrained embeddings using unlabeled gallery data. The method has two stages: 1) refine gallery features through multi-scale neighborhood diffusion, 2) online query adaptation using the refined gallery and anchored score diffusion. Experiments show consistent improvements on different pretrained backbones as well as generalization to other data types. However, there are critical concerns about the technical contributions, robustness and scalability of the proposed method.

Critical strengths
  1. Simple method with strong experimental results: the proposed method is pretty simple and straightforward but demonstrates strong results across different backbones as well as generalizability to other data types such as point cloud and sketch-based data.
  2. Efficiency: GEAR processes and adapts each query independently using the precomputed gallery without the need for training, which is useful for online retrieval, especially when labeled data is not available.
Critical weaknesses
  1. Limited novelty and unclear component contribution: the framework mainly combines existing techniques (centering, neighborhood aggregation, graph diffusion) with no new technical advancements. Also, the ablations do not justify the proposed components: Tab 7 shows multi-K refinement has similar results as single-K, Tab 9 shows ordinary diffusion achieves better mAP and ANMRR than GEAR, which makes it even harder to justify the contribution of the paper
  2. Limited robustness evaluation and scalability: the quality of the initial gallery neighborhoods heavily affects GEAR, which can be a problem when we have highly similar objects from different classes in the gallery. Also, the offline stage seems to have quadratic complexity which can be hard to scale when the size of the gallery increases.
Questions for authors

Please see the weaknesses and address the concerns.

10 comments
AuthorsOpenReview ↗

Response to Reviewer eiTv [1/5]

Response to Weakness 1[1/3]

Thank you for your thorough and constructive review. For your Critical Weaknesses 1, we sincerely respond as follows:

Regarding novelty: We propose a new training-free paradigm for open-set 3D retrieval through test-time diffusion-based adaptation

First, GEAR offers a practical perspective on open-set 3D retrieval: exploiting an unlabeled gallery to improve retrieval for independently arriving queries, without costly task-specific training, target-domain labels, or gradient-based test-time optimization. Unlike existing training-based approaches, GEAR treats the available gallery itself as a reusable source of adaptation. Once the gallery is prepared offline, each incoming query can be processed independently, without access to other queries or further optimization. This shifts the focus from learning or optimizing representations using additional supervision or query sets to exploiting the intrinsic structure of the unlabeled retrieval gallery.

Second, our method is not a simple combination of established techniques. We further conducted two sets of controlled experiments: comparisons against classical retrieval pipelines and a query-module replacement study. The results show that GEAR outperforms the evaluated classical combinations, while its query recalibration provides additional gains even on the same classically DBA-refined gallery.

1. Comparison with classical retrieval pipelines. We compare GEAR against representative classical methods and their combination using identical frozen descriptors, splits, and evaluation protocols. Classical hyperparameters are selected through leave-one-dataset-out (LODO) validation, while GEAR retains its submitted configuration.

Method (Macro mAP, %)CLIPDINOv2DINOv3
Frozen descriptors56.3860.9860.64
Gallery centering only60.2163.8363.15
Weighted AQE + DBA63.7566.5965.99
Standard global diffusion65.3666.9966.05
Classical combination*63.9966.2765.20
GEAR (Ours)65.6568.5367.96

Classical combination: Gallery centering + weighted AQE/DBA + restart-based graph propagation.

GEAR outperforms this classical combination by +1.66/+2.26/+2.76 mAP points on CLIP/DINOv2/DINOv3, respectively, with positive paired 95% confidence intervals.

2. Query recalibration improves retrieval on the same classical DBA gallery.

To isolate the contribution of query recalibration, we use the same final gallery refined by classical iterative weighted DBA for both methods.

Classical Query performs two rounds of weighted AQE along the classical gallery refinement trajectory, with LODO-selected parameters, whereas GEAR Query applies one-step recalibration against the same final gallery. Both share identical frozen descriptors, gallery centering, and evaluation settings, without GEAR's gallery refinement or score diffusion.

BackboneMethodmAP (%)ΔmAP (pp, paired 95% CI)
CLIPClassical Query65.06—
GEAR Query (Ours)65.61+0.55 [0.01, 1.14]
DINOv2Classical Query68.05—
GEAR Query (Ours)68.49+0.44 [0.01, 0.89]
DINOv3Classical Query67.35—
GEAR Query (Ours)67.90+0.55 [0.05, 1.03]

GEAR Query improves macro mAP by 0.44–0.55 points across all three backbones, with positive paired confidence intervals. This demonstrates that GEAR's query recalibration provides measurable gains even when using a conventional DBA-refined gallery.

Together, the classical-pipeline comparison and this controlled replacement study show that GEAR's gains are not reproduced by the evaluated classical combination, while its query recalibration makes an additional contribution under matched gallery representations.

AuthorsOpenReview ↗

Response to Reviewer eiTv [2/5]

Response to Weakness 1[2/3]

Regarding Multi-K: New density and per-dataset experiments support a competitive fixed configuration

We agree that Multi-K and Single-K (K=20K=20) achieve similar macro mAP in Table 7. Our motivation for Multi-K is not to outperform every single scale, but to accommodate heterogeneous neighborhood structures without selecting a dataset-specific K using target labels. To directly examine this motivation, we conducted additional density-stratified and per-dataset experiments across three backbones and four datasets.

1. Neighborhood-scale heterogeneity. We group gallery classes by local density, measured using the mean 10-NN similarity of centered frozen descriptors, and evaluate queries from each group without changing the gallery. Class labels are used only for post-hoc analysis. The best Single-K differs between low- and high-density groups in 8/12 backbone–dataset settings. For example, on DINOv2/ABO:

Density groupBest Single-KSingle-K mAPMulti-K mAPΔmAP (pp)
Low1056.7156.51−0.20
Medium1074.5474.65+0.11
High5084.1383.95−0.18

ΔmAP = Multi-K − post-hoc best Single-K.

The best Single-K differs across density groups within the same gallery, while Multi-K achieves competitive performance without selecting K separately for each group. Across all 12 medium-density groups, Multi-K stays within 0.70 mAP points of the post-hoc best Single-K (average gap: −0.41 points).

2. Per-dataset Single-K comparison. We further evaluate each neighborhood scale on individual datasets to examine whether a single K is consistently preferable. The DINOv2 results are shown below.

DatasetK=5K=10K=20K=50Multi-K
ESB63.2463.5563.4863.1263.31
NTU65.2965.6065.3064.7865.44
MN4073.0973.7774.4175.1074.40
ABO70.5271.0370.9870.3770.99

Values are mAP (%). Bold indicates the best Single-K result for each dataset.

The preferred Single-K varies across datasets: K=10 performs best on ESB, NTU, and ABO, whereas K=50 performs best on MN40. Multi-K achieves competitive performance across all four datasets without selecting K separately for each target dataset. These results support using a fixed multi-scale configuration to accommodate heterogeneous neighborhood structures, although Multi-K does not outperform the best Single-K on every dataset.

3. Accuracy–efficiency trade-off. We additionally measure the complete offline and online costs using matched implementations. On ModelNet40 (9,329 gallery objects, DINOv2):

MethodmAP (%)Offline preparationOnline P50/queryFinal cache
Single-K (K=20K=20)74.4112.56 s4.99 ms29.28 MiB
Multi-K74.4058.07 s4.90 ms29.27 MiB

Multi-K incurs additional one-time offline computation, but its fused representation is cached, resulting in nearly unchanged online latency and cache size. We therefore present Multi-K as a fixed multi-scale design for heterogeneous gallery neighborhoods.

Overall, these experiments clarify the contribution of Multi-K: its purpose is to accommodate varying neighborhood-scale preferences through a unified, label-free fusion strategy, rather than to guarantee the highest mAP at every individual scale. The chosen K∈{5,10,20,50}K\in\{5,10,20,50\} spans local to broader neighborhoods, with equal-weight fusion avoiding additional target-specific weight selection. We will incorporate these analyses in the revision.

AuthorsOpenReview ↗

Response to Reviewer eiTv [3/5]

Response to Weakness 1[3/3]

Regarding anchored diffusion: New experiments show better top-rank accuracy and graph robustness

We agree that ordinary diffusion achieves slightly higher mAP and better ANMRR than anchored diffusion in Table 9. However, anchored diffusion aims to preserve reliable query evidence during propagation, rather than maximize mAP alone. Our additional experiments show that it substantially improves Top-1 retrieval accuracy and robustness to erroneous graph connections.

1. Top-rank accuracy at the submitted L=2L=2. We compare ordinary and anchored diffusion using identical initial query scores s0s_0, gallery graphs PP, and propagation depths across three backbones and four datasets.

BackboneMethodmAP (%)P@1 (%)NDCG@100 gain
CLIPOrdinary65.8578.79—
Anchored (Ours)65.6584.60+0.23
DINOv2Ordinary68.7482.46—
Anchored (Ours)68.5386.75+0.17
DINOv3Ordinary68.0782.47—
Anchored (Ours)67.9687.95+0.15

Anchoring improves P@1 by 4.30–5.81 percentage points over ordinary diffusion, with positive paired 95% confidence intervals, at a small mAP cost (0.11–0.21 points). It also retains a 0.35–0.38-point mAP gain over no diffusion.

2. Robustness to erroneous graph connections. To test whether anchoring actually limits harmful propagation, we keep gallery features, initial scores, and L=2L=2 fixed, and perturb only the final gallery graph. We evaluate both random edge rewiring and wrong-class edge rewiring using three fixed seeds.

BackboneRandom rewiring (25%)Wrong-class rewiring (25%)
CLIP7.287.93
DINOv27.177.86
DINOv37.177.79

Values indicate the reduction in mAP degradation (percentage points) achieved by anchoring relative to ordinary diffusion; paired 95% confidence intervals are positive in all six conditions.

These results directly support the intended role of anchoring: when gallery connections become unreliable, retaining the initial query evidence substantially reduces the damage caused by score propagation. The wrong-class rewiring is a controlled diagnostic; class labels are not used during normal inference.

3. Stability across propagation depths. We additionally evaluate L∈{1,2,5,10,20}L\in\{1,2,5,10,20\} in 12 backbone–dataset settings. At the submitted fixed L=2L=2, the average gap to each method's own post-hoc best depth is only 0.015 mAP points for anchoring, compared with 0.618 for ordinary diffusion. This indicates that anchoring is less sensitive to propagation depth in the tested settings, consistent with Figure 3.

Overall, we retain anchored diffusion because it provides stronger top-rank accuracy, greater resistance to controlled graph corruption, and lower depth sensitivity: despite slightly lower clean mAP and ANMRR than ordinary diffusion. These additional results clarify its contribution beyond the marginal mAP improvement reported in Table 9.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

As the discussion period draws to a close, we would like to express our sincere gratitude to you for the time and effort dedicated to reviewing our manuscript. We deeply appreciate the thoughtful and constructive feedback.

We are greatly encouraged to see that our reviewers recognize GEAR as a practical, effective, and flexible framework for test-time 3D object retrieval.

In particular, following our detailed rebuttal and new experiments, Reviewers eiTv and L38M have confirmed that their primary concerns have been addressed and have expressed positive support for accepting our work.

Furthermore, although Reviewer MFFi has not yet replied to our rebuttal, we deeply appreciate his/her positive support in the initial evaluation for accepting our work. We hope that our point-by-point responses and revisions further strengthen the confidence in supporting our work.

  • Reviewers L38M, MFFi, and eiTv recognize that our training-free, transductive formulation is highly practical, as it bypasses the need for costly 3D annotations and domain retraining.
  • Reviewers L38M and eiTv find that our asymmetric offline-online design elegantly resolves the batch-query dependency of conventional transductive methods, making it highly applicable to standard online single-query scenarios.
  • Reviewers MFFi and L38M appreciate the comprehensive experimental results, noting that our framework serves as a versatile, plug-and-play module that consistently boosts diverse pretrained representations.

For clarity and to assist the Area Chair's final assessment, we briefly recap the core contributions of our framework:

  • Formulating test-time 3D object retrieval adaptation as a training-free transductive problem, which exploits intrinsic neighborhood relationships within unlabeled galleries to improve embedding discriminability without requiring downstream annotations.
  • An Asymmetric Offline-Online Framework that enables independent single-query adaptation through reusable gallery representations, successfully eliminating gradient optimization and batch-query dependencies.
  • An Iterative Anchored Diffusion strategy that refines retrieval rankings while robustly preserving original query evidence through repeated score anchoring.

Guided by the reviewers' valuable suggestions, we have conducted extensive new experiments. The manuscript has been carefully revised, with all major updates highlighted in blue.

New Experiments & Additional Results:

  • Classical Baselines: We compared GEAR with classical retrieval methods (QE, DBA, conventional graph diffusion, and their combinations) and conducted controlled query-recalibration experiments (Appendix D.6).
  • DEC Plug-and-Play Evaluations: We provided backbone-matched comparisons with DEC (Appendix C.2), obtaining additional macro mAP gains of 6.45 and 6.48 points over DEC alone on DINOv2 and DINOv3.
  • Large-Gallery Scalability: We evaluated matched Single-K vs. Multi-K efficiency (Appendix E.3) and scaled our evaluation to a 50,000-object gallery using exact and approximate (HNSW) nearest-neighbor search (Appendix E.5).
  • Multi-K Justification: We included per-dataset and density-stratified neighborhood analyses, cross-dataset parameter selection, and alternative fusion strategies to substantiate the fixed Multi-K configuration (Appendix D.2).
  • Diffusion Robustness: We conducted top-rank preservation, graph perturbation, and propagation-depth stability experiments to clarify the accuracy-robustness trade-offs of anchored diffusion (Appendix D.3 & D.7).

Other Modifications & Clarifications:

  • Theoretical Grounding: We clarified the statistical interpretation of gallery centering (Section 3.3.1), supported by theoretical references and cross-class consistency validation (Appendix F.4).
  • Related Work: We expanded Section 2 to discuss recent related studies suggested by the reviewers, including Source-Free Model Adaptation, Uni-Adapter, and PASR.
  • Structural Diagnostics: We added comprehensive gallery structure diagnostics (neighborhood purity, Fisher ratio, and hubness skewness) to structurally analyze the embedding space (Appendices F.1–F.3).

We hope this summary highlights the rigorous updates made to the manuscript. We sincerely appreciate your time, support, and guidance throughout this review process.

Best regards,

The Authors

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

Related papers

Discussion (0)

Sign in to comment.