Pith. sign in

REVIEW 3 major objections 6 minor 41 references

For each 3D shape query, GATE-3D predicts how strongly geometry should correct the appearance-based ranking from cross-modal disagreement features, improving open-set retrieval on geometry-sensitive domains while leaving appearance-dominate

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A per-query router trained on cross-modal disagreement features blends a geometry residual into appearance-based 3D-shape retrieval, improving mAP@10 by 2.0 points on OS-ESB-core while avoiding always-on fusion degradation.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A careful, honest paper on a modest idea whose central geometry signal needs external validation before the results can be trusted. the 3 major comments →

arxiv 2607.19111 v1 pith:CZ6FDTTR submitted 2026-07-21 cs.CV

GATE-3D: Geometry-Aware Test-time Adaptive Reranking for Open-Set 3D Shape Retrieval

Hao Wu , Heyi Lin , Zilin Wang , Huizai Yao , Hao Wang , Hui Xiong This is my paper
classification cs.CV
keywords 3D shape retrievalopen-set retrievalcross-modal disagreementquery-adaptive rerankinggeometry-aware featurestest-time adaptationlinear routinggeometric false positives
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

GATE-3D tries to establish that geometry should not be fused with appearance unconditionally in 3D shape retrieval: a per-query decision about how much geometric correction to apply beats both appearance-only retrieval and always-on fusion. The decision is made by a lightweight linear router that reads 24 cross-modal disagreement features—how much the appearance-based and geometry-based rankings of the retrieved candidates diverge—and outputs a blending weight. On a geometry-sensitive mechanical-part benchmark, this improves mAP@10 by 2.00 points over appearance-only retrieval, reduces geometric false positives by 10.8%, and preserves gains on unseen categories; on appearance-dominated benchmarks it defaults to near-appearance-only behavior and avoids the degradation that unconditional fusion causes. The paper also argues that in the low-data regime, well-designed disagreement features matter more than model capacity, since a linear model beats a small MLP.

Core claim

GATE-3D's central claim is that cross-modal disagreement between appearance and geometry rankings is a reliable, measurable signal for when geometry should correct an appearance-based retrieval. Concretely, the paper defines a geometry residual ΔS_geo = S_geoaware − S_app for each gallery item and lets a per-query weight α_q (predicted by Ridge or logistic regression from a 24-dimensional disagreement feature vector) control how much of that residual is added: S_final = S_app + α_q·ΔS_geo. The paper shows that this query-adaptive residual reranking, trained under strict out-of-fold cross-validation, improves open-set mechanical part retrieval (mAP@10 +2.00 pp, p=0.041), reduces geometric fal

What carries the argument

The load-bearing mechanism is the 24-dimensional cross-modal disagreement feature vector φ(q), computed over the top-50 candidates retrieved by appearance. It contains rank correlations (Kendall-τ and Spearman-ρ between appearance and each geometry modality's ranking), score-distribution statistics (mean absolute deviations of score differences and top-10 rank-inversion fractions), and margin features (top-1 vs top-5 score gaps per modality). These features feed a lightweight linear router (Ridge regression for continuous α_q, logistic regression for binary gating) that predicts the blending weight in the residual score S_final = S_app + α_q·ΔS_geo. The geometry residual itself, ΔS_geo, isol

Load-bearing premise

The paper assumes that a randomly initialized depth-map encoder and a point-cloud histogram genuinely capture the geometric structure that appearance misses; if that signal is mostly noise, the router is learning to route on noise and the measured gains would not reflect genuine geometric understanding.

What would settle it

Replace the randomly initialized depth encoder with an ImageNet-initialized version (or any fixed alternative) and re-run the mechanical-part benchmark: if the claimed +5.9 pp gain from random initialization disappears, the geometry signal is an artifact of that specific initialization, and the routing gains are not evidence of geometric reasoning.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Any appearance-based retrieval pipeline can gain a query-adaptive geometry stage as a drop-in post-hoc reranker, with no retraining of the underlying feature extractors.
  • On geometry-sensitive domains such as mechanical part search, the method improves mAP@10 by 2.00 points over appearance-only retrieval and cuts geometric false positives by about 10.8%.
  • On appearance-dominated domains, the router activates geometry for only a small fraction of queries (about 5%), avoiding the degradation that always-on fusion causes.
  • In the low-data regime, a simple linear router outperforms a small neural network, implying that the quality of disagreement features, not model capacity, drives routing success.
  • Leave-one-category-out evaluation shows the routing signal transfers to unseen shape subcategories, supporting open-set deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The residual-gating design (base score + weighted correction) generalizes beyond shape retrieval: any pair of modalities where one is strong but blind to a specific distinction could use disagreement features to gate a correction, as the paper itself hints for medical and satellite imagery.
  • The result that cross-modal disagreement predicts per-query geometric utility suggests a broader design principle for multi-modal systems: measuring how modalities diverge can be more informative than measuring each modality's confidence alone.
  • The paper reports that a gate trained on mechanical parts does not transfer to a different benchmark, suggesting the disagreement features are domain-specific; a natural extension is to learn domain-agnostic disagreement statistics or adapt the gate with a handful of labeled queries in the target domain.
  • With the oracle upper bound about 8 points above the method, there is headroom; testing per-modality residuals (depth, point cloud, normal gated separately) is a concrete way to capture more of that gap, as the paper mentions as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes GATE-3D, a query-adaptive reranking method for open-set 3D shape retrieval. Appearance-based retrieval is performed with a frozen DINOv2 (or CLIP) backbone; a geometry-aware score is built from randomly initialized ResNet18 depth features, a PCA+histogram point-cloud descriptor, and optional surface normals. The method computes 24 cross-modal disagreement features from the top-K appearance and geometry rankings and trains a lightweight router (Ridge regression or logistic regression) to predict a per-query blending weight alpha_q. The final score is S_final = S_app + alpha_q * Delta_S_geo (Eq. 4), where Delta_S_geo is the difference between a geometry-aware aggregate and the appearance score. Training and evaluation use an out-of-fold protocol, with leave-one-category-out evaluation. Experiments on OS-ESB-core, OS-NTU-core, and OS-MN40-core claim a +2.00 pp mAP@10 improvement over appearance-only retrieval on OS-ESB-core (one-sided p=0.041), a 10.8% reduction in geometric false positives, and more robust behavior than always-on fusion on appearance-dominated benchmarks.

Significance. If the central claims hold, GATE-3D is a useful, backbone-agnostic reranking layer that activates geometry only when cross-modal disagreement indicates it will help. The paper has notable strengths: the out-of-fold protocol and leave-one-category-out evaluation are credible, the ablations control for query-only features and generic retrieval ambiguity, and the candidate-pool sensitivity analysis supports the safety claim. The finding that simple linear routing beats a small MLP in the low-data regime is interesting and well documented. However, the primary improvement relies on a one-sided p-value on a single metric (mAP@10), and the geometry signal itself rests on a randomly initialized depth encoder whose geometric validity is not externally validated. The safety advantage over always-on fusion is well supported, but the significance of the claimed improvement over appearance-only retrieval is more fragile.

major comments (3)
  1. [§3.1, Eq. (3)] The depth channel is the load-bearing component of Delta_S_geo, yet it is a randomly initialized ResNet18 with no training. The only evidence for its geometric utility is a stated +5.9 pp mAP improvement over ImageNet-init on OS-ESB-core, reported without experimental detail, error bars, or qualitative verification. Because Eq. (4) applies this residual as the only geometry correction, the framework's 'geometry awareness' is unsubstantiated if the random projections encode appearance artifacts rather than geometric structure. Please provide external validation: qualitative nearest-neighbor examples for geometrically different but visually similar parts (e.g., blind vs through holes), a comparison against a supervised/self-supervised depth encoder, or controlled synthetic experiments showing that the random-init features separate geometry that appearance alone cannot.
  2. [§4.2, §5, Table 1] The central claim of improvement over appearance-only retrieval rests on a one-sided p=0.041 on mAP@10, with a two-sided 95% CI that includes zero, as the paper itself acknowledges. Moreover, on the primary standard mAP, GATE-3D-alpha (58.58) does not beat always-on fusion (58.65), and GATE-3D LR (57.88) is only +0.86 over DINOv2. The safety claim is well supported, but the improvement claim is overstated in the abstract. Please report the two-sided p-value and CI for the primary metric, and either pre-specify the primary metric or correct for the multiple metrics/variants tested. If the paper's intended contribution is robustness rather than raw improvement, that should be the headline claim.
  3. [§3.5, Eq. (2)] The geometry aggregation weights (alpha_app, alpha_dep, alpha_nor, alpha_geo) are determined by Dirichlet-sampled nested grid search on training queries. It is not clear whether the +5.9 pp random-init advantage in §3.1 was obtained with this same weight-fitting procedure, nor how stable the selected weights are across outer folds. With only 120 training queries and 22 positives, fitting these aggregation weights plus the router hyperparameters is a nontrivial source of overfitting. Please report the selected weights per fold, the number of Dirichlet samples, and how the inner/outer split prevents leakage of the weight search into evaluation. Otherwise the reader cannot separate the contribution of the geometry signal from the contribution of hyperparameter selection.
minor comments (6)
  1. [Abstract, Fig. 2] Typos: 'rEranker' in the abstract, 'Apperance' in Figure 2, and 'OS-ESB-corecontains' missing space. Also inconsistent 'HGM2R' vs 'HGM 2R' spacing.
  2. [§3.1] The point-cloud descriptor is described as 'PCA + histogram descriptor (33-dim) capturing spatial extent and local curvature [37]' but no precise definition is given. Please specify the exact computation or cite a concrete algorithm.
  3. [Table 1, caption] The 'Mod.' column abbreviations (I., G., P., V., etc.) are not defined in the caption. Add a legend for readability.
  4. [§3.3, Eq. (7)] The indicator 1[g(q)>0.5] in Eq. (7) is standard but the notation is not introduced; define it as the 0/1 indicator.
  5. [§4.3, Table 2] The 'In-sample reference (upper bounds only)' block contains 'GATE-3D MLP, binary routing 0.8357 +3.19', which is not an upper bound but a trained in-sample model. This row should be removed from the upper-bound list or clearly relabeled as in-sample only, since it does not represent a valid generalization estimate.
  6. [References] Citation placement in the introduction is malformed: '[1, 9, 11] Self-supervised models...' places the bracket before the sentence subject. Also, some references appear as '[31–34]' after a sentence without textual integration; please fix the citation style.

Circularity Check

0 steps flagged

No significant circularity: the router is trained out-of-fold on an oracle target, and the disagreement features are inputs rather than fitted outputs.

full rationale

GATE-3D's central claim is that a per-query blending weight alpha_q can be predicted from 24 cross-modal disagreement features and applied as a residual correction (Eq. 4). This is not circular: the prediction target (Eq. 6 oracle alpha) is computed on training-fold queries only, and the router is evaluated on held-out folds under a 5-fold x 5-seed OOF protocol. The disagreement features are statistics of the appearance and geometry rankings (Kendall/Spearman correlations, score deviations, margins); they are inputs to the router, not equal to alpha or to the AP gain, so the mapping from features to alpha is learned rather than definitional. The geometry-fusion weights in Eq. 2 are selected by nested CV on training queries, which is legitimate hyperparameter fitting; the oracle target's use of AP is standard supervised regression, not a leak, because held-out evaluation uses the predicted alpha, not the oracle. The random-init depth encoder and the point-cloud descriptor are design choices supported by an internal ablation, and the Limitations section concedes cases where geometry cues are confounded, but that is an empirical limitation, not a circular derivation. No load-bearing self-citations appear: references to HGM2R, DAC, TeDA, and Open3D are external baselines or tools. The GFP reduction is closely aligned with the mechanism, but it is an auxiliary empirical metric rather than the derivation of the main result. Overall, the derivation is self-contained with respect to its claimed prediction.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central claim rests on several fitted hyperparameters (geometry fusion weights, router regularization, K, oracle grid, ensemble weight) and on assumptions about DINOv2 features, random-initialized depth features, benchmark labels, and statistical testing. No new physical entities are introduced.

free parameters (7)
  • geometry fusion weights (alpha_app, alpha_dep, alpha_nor, alpha_geo) = alpha_nor=0.014 reported; others not stated
    Eq. 2 combines modalities; weights determined by Dirichlet-sampled nested grid search on training queries (Section 3.1). They define Delta_S_geo and hence the whole reranking signal.
  • Ridge L2 strength lambda (GATE-3D-alpha) = not stated
    Tuned by inner 3-fold CV (Section 3.3); controls shrinkage of the continuous router.
  • Logistic regression C (GATE-3D LR) = C=0.1
    L2 regularization for the LR router (Section 3.4); chosen for the low-data regime.
  • Fallback threshold epsilon = 0.15
    Clips small predicted alpha to 0 (Section 3.3); a hand-set hyperparameter affecting activation rate.
  • Candidate pool size K = 50
    Selected by inner-fold mAP@10 over K in {20,30,50,100} (Section 3.5).
  • Oracle alpha grid = alpha in {0,0.1,...,1.0}
    Training target grid in Eq. 6; the authors verify finer grids give <0.1 pp difference, but the grid choice is a free design decision.
  • Ensemble weight w (DINOv2+CLIP) = not stated
    Eq. 8 convex ensemble; selected by inner-fold CV (Section 3.6); used for App. Fusion results.
axioms (5)
  • domain assumption DINOv2 pretrained features are a strong appearance representation for 3D shapes when aggregated over 12 rendered views.
    Section 3.1 uses DINOv2-ViT-B/14 max-pooled over views as the appearance backbone; the paper relies on prior results rather than validating this choice.
  • ad hoc to paper A randomly initialized ResNet18 on depth maps and a PCA+histogram point-cloud descriptor encode geometric structure relevant to retrieval.
    Section 3.1 claims random-init depth features yield +5.9 pp mAP over ImageNet init on OS-ESB-core; this surprising claim is supported only by internal ablation, with no external evidence that random projections capture blind-vs-through-hole geometry.
  • domain assumption The per-query oracle alpha* (Eq. 6), computed by grid search maximizing AP over the candidate pool with ground-truth labels, is a learnable and stable target.
    Section 3.3 defines the regression target; if the 120-query/22-positive training set is unrepresentative, the router cannot learn a transferable policy.
  • domain assumption Ground-truth relevance labels and open-set splits of OS-ESB-core, OS-NTU-core, and OS-MN40-core are correct and comparable to the DAC/HGM2R protocols.
    All evaluation rests on these benchmarks (Section 4.1); label errors or split leakage would change results.
  • domain assumption One-sided paired t-test over OOF folds is a valid significance test for the mAP@10 improvement.
    Section 4.2/5 report p=0.041 one-sided with two-sided CI [-0.0001,+0.0444]; treating folds as independent samples is standard but the small n makes inference fragile.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GATE-3D: Geometry-Aware Test-time Adaptive Reranking for Open-Set 3D Shape Retrieval." pith.science (2026). https://pith.science/paper/CZ6FDTTR

@misc{pith2026260719111,
  author       = {Pith},
  title        = {Pith review of: GATE-3D: Geometry-Aware Test-time Adaptive Reranking for Open-Set 3D Shape Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZ6FDTTR}},
  note         = {Machine review of arXiv:2607.19111}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large pretrained vision models have substantially improved appearance-based 3D shape retrieval, but they still confuse shapes that look similar while differing in geometry. Although geometry-aware features can reduce these errors, naive fusion of geometry and appearance may hurt retrieval when the two modalities are already well aligned. We propose GATE-3D, a lightweight query-adaptive reranking method that incorporates geometry without retraining the retrieval backbone. For each query, GATE-3D predicts how much a geometry-aware score should adjust the appearance-based ranking using features that capture disagreement between the two modalities. This selective design lets geometry contribute where it helps and stay silent where it would hurt. Experiments on three open-set 3D retrieval benchmarks show that GATE-3D improves over appearance-only retrieval and is more robust than always-on fusion. On the primary benchmark, it improves mAP@10 by 2.00 points over appearance-only retrieval (p=0.041); it also improves leave-one-category-out generalization and reduces geometric false positives by 10.8%. GATE-3D achieves competitive zero-shot results against DAC-based baselines. We further find that simple linear routing is more effective than a small MLP in the low-data regime, suggesting that cross-modal disagreement features matter more than model capacity for adaptive routing.

Figures

Figures reproduced from arXiv: 2607.19111 by Hao Wang, Hao Wu, Heyi Lin, Hui Xiong, Huizai Yao, Zilin Wang.

Figure 1
Figure 1. Figure 1: Motivation for query-adaptive geometry routing. Given a mechanical query part, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the GATE-3D framework. Left (Offline Feature Extraction): each 3D shape is rendered into 12 views and encoded by DINOv2 (appearance), a randomly initialized ResNet18 (depth), and a PCA+histogram descriptor (point cloud); surface normals are optionally included. A geometry residual Δ𝑆geo is computed as the difference between the geometry-aware aggregate 𝑆geoaware and the appearance score 𝑆app. R… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative retrieval comparison on OS-ESB-core. Top-5 retrieved shapes for a flange-type query under DAC (CLIP), [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 4 linked inside Pith

  1. [1]

    Song Bai, Xiang Bai, Zhichao Zhou, Zhaoxiang Zhang, and Longin Jan Latecki

  2. [2]

    Ding-Yun Chen, Xiao-Pei Tian, Yu-Te Shen, and Ming Ouhyoung. 2003. On visual similarity based 3D model retrieval. InComputer graphics forum, Vol. 22. Wiley Online Library, 223–232

    2003

  3. [3]

    Zhimin Chen, Xuewei Chen, Xiao Guo, Yingwei Li, Longlong Jing, Liang Yang, and Bing Li. 2025. Point Cloud Self-supervised Learning via 3D to Multi-view Masked Learner. InProceedings of the IEEE/CVF International Conference on Computer Vision. 27618–27629

    2025

  4. [4]

    Leszek Cieplinski. 2001. MPEG-7 color descriptors and their applications. In International Conference on Computer Analysis of Images and Patterns. Springer, 11–20

    2001

  5. [5]

    Yifan Feng, Shuyi Ji, Yu-Shen Liu, Shaoyi Du, Qionghai Dai, and Yue Gao. 2023. Hypergraph-based multi-modal representation for open-set 3D object retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence46, 4 (2023), 2206– 2223

    2023

  6. [6]

    Rao Fu, Jie Yang, Jiawei Sun, Fang-Lue Zhang, Yu-Kun Lai, and Lin Gao. 2020. RISA-Net: Rotation-invariant structure-aware network for fine-grained 3D shape retrieval.arXiv preprint arXiv:2010.00973(2020)

    Pith/arXiv arXiv 2020

  7. [7]

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. InInternational conference on machine learning. PMLR, 1321–1330

    2017

  8. [8]

    Junjie Guo, Chenqiang Gao, Fangcen Liu, Deyu Meng, and Xinbo Gao. 2024. Damsdet: Dynamic adaptive multispectral detection transformer with compet- itive query selection and adaptive feature fusion. InEuropean Conference on Computer Vision. Springer, 464–481

    2024

  9. [9]

    Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. 2021. Mvtn: Multi-view transformation network for 3d shape recognition. InProceedings of the IEEE/CVF international conference on computer vision. 1–11

    2021

  10. [10]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition. 770–778

    2016

  11. [11]

    Xinwei He, Yang Zhou, Zhichao Zhou, Song Bai, and Xiang Bai. 2018. Triplet- center loss for multi-view 3d object retrieval. InProceedings of the IEEE conference on computer vision and pattern recognition. 1945–1954

    2018

  12. [12]

    Subramaniam Jayanti, Yagnanarayanan Kalyanaraman, Natraj Iyer, and Karthik Ramani. 2006. Developing an engineering shape benchmark for CAD models. Computer-Aided Design38, 9 (2006), 939–953

    2006

  13. [13]

    Maurice G Kendall. 1938. A new measure of rank correlation.Biometrika30, 1-2 (1938), 81–93

    1938

  14. [14]

    Jian Liang, Ran He, and Tieniu Tan. 2025. A comprehensive survey on test-time adaptation under distribution shifts.International Journal of Computer Vision 133, 1 (2025), 31–64

    2025

  15. [15]

    Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. 2023. Openshape: Scaling up 3d shape representation towards open-world understanding.Advances in neural information processing systems36 (2023), 44860–44879

    2023

  16. [16]

    Daniel Maturana and Sebastian Scherer. 2015. Voxnet: A 3d convolutional neural network for real-time object recognition. In2015 IEEE/RSJ international conference on intelligent robots and systems (IROS). Ieee, 922–928

    2015

  17. [17]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193(2023)

    Pith/arXiv arXiv 2023

  18. [18]

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. 2017. Pointnet: Deep learning on point sets for 3d classification and segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition. 652–660

    2017

  19. [19]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al

  20. [20]

    Charles Spearman. 1987. The proof and measurement of association between two things.The American journal of psychology100, 3/4 (1987), 441–471

    1987

  21. [21]

    Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. 2015. Multi-view convolutional neural networks for 3d shape recognition. InProceed- ings of the IEEE international conference on computer vision. 945–953

    2015

  22. [22]

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. 2020. Test-time training with self-supervision for generalization under distribution shifts. InInternational conference on machine learning. PMLR, 9229– 9248

    2020

  23. [23]

    Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. 2022. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. Advances in neural information processing systems35 (2022), 14388–14402

    2022

  24. [24]

    Zhichuan Wang, Yang Zhou, Zhe Liu, Rui Yu, Song Bai, Yulong Wang, Xinwei He, and Xiang Bai. 2025. Describe, Adapt and Combine: Empowering CLIP Encoders for Open-set 3D Object Retrieval. InProceedings of the IEEE/CVF International Conference on Computer Vision. 21026–21036

    2025

  25. [25]

    Zhichuan Wang, Yang Zhou, Jinhai Xiang, Yulong Wang, and Xinwei He. 2025. Teda: Boosting vision-lanuage models for zero-shot 3d object retrieval via testing- time distribution alignment. InProceedings of the 2025 International Conference on Multimedia Retrieval. 1442–1451

    2025

  26. [26]

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao. 2015. 3d shapenets: A deep representation for volumet- ric shapes. InProceedings of the IEEE conference on computer vision and pattern recognition. 1912–1920

    2015

  27. [27]

    Yang Xu, Yifan Feng, Jun Zhang, Jun-Hai Yong, and Yue Gao. 2024. Assembly fuzzy representation on hypergraph for open-set 3D object retrieval.Advances in Neural Information Processing Systems37 (2024), 125098–125115

    2024

  28. [28]

    Yang Xu, Yifan Feng, Jun Zhang, Jun-Hai Yong, and Yue Gao. 2024. Semi-open 3D object retrieval via hierarchical equilibrium on hypergraph.Advances in Neural Information Processing Systems37 (2024), 63250–63267

    2024

  29. [29]

    Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martín- Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. 2024. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 27091– 27101

    2024

  30. [30]

    Zihui Xue and Radu Marculescu. 2023. Dynamic multimodal fusion. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2575–2584

    2023

  31. [31]

    Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu

  32. [32]

    Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. 2022. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training.Advances in neural information processing systems35 (2022), 27061–27074

    2022

  33. [33]

    Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. 2022. Pointclip: Point cloud understanding by clip. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8552–8562

    2022

  34. [34]

    Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. 2023. Learn- ing 3d representations from 2d pre-trained models via image-to-point masked autoencoders. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 21769–21780

    2023

  35. [35]

    Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. 2017. Re-ranking person re-identification with k-reciprocal encoding. InProceedings of the IEEE conference on computer vision and pattern recognition. 1318–1327

    2017

  36. [36]

    Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. 2023. Uni3d: Exploring unified 3d representation at scale.arXiv preprint arXiv:2310.06773(2023)

    Pith/arXiv arXiv 2023

  37. [37]

    Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. 2018. Open3D: A modern library for 3D data processing.arXiv preprint arXiv:1801.09847(2018)

    Pith/arXiv arXiv 2018

  38. [38]

    Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. 2023. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. InProceedings of the IEEE/CVF international conference on computer vision. 2639–2650

    2023

  39. [2016]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    Gift: A real-time and scalable 3d shape search engine. InProceedings of the IEEE conference on computer vision and pattern recognition. 5023–5032

  40. [2021]

    In International conference on machine learning

    Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR, 8748–8763

  41. [2022]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 19313–19322

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.