pith. sign in

arxiv: 2606.23069 · v2 · pith:WUCS3TF6new · submitted 2026-06-22 · 💻 cs.CV

Rethinking Prototype-based Similarity Learning for Few-Shot Object Detection

KunHo Heo , Seungjae kim , Wongyu Lee , SuYeon Kim , MyeongAh Cho This is my paper

Pith reviewed 2026-06-30 10:35 UTC · model grok-4.3

classification 💻 cs.CV
keywords few-shot object detectionprototype-based similarity learningtext-anchored semantic maskstage-aligned hierarchical autoregressive regressioninter-class similarity marginViT layer alignmentCOCO benchmark
0
0 comments X

The pith

Text-anchored semantic masks and stage-aligned regression resolve inter-class confusion and localization issues in prototype-based few-shot object detection.

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

The paper identifies two core problems in prototype-based similarity learning for few-shot object detection: similarity scores cause classes to overlap in feature space and supply too little spatial detail for accurate boxes. It proposes Text-Anchored Semantic Mask to use text features as anchors that pick relevant visual channels and widen the gaps between classes, plus Stage-Aligned Hierarchical Autoregressive Regression that refines boxes step by step by matching deeper ViT layers to coarse localization and shallower layers to fine spatial cues. If these components work as described, prototype methods can adapt to new classes without retraining while producing both correct labels and tight boxes. Experiments on the COCO benchmark show the combined changes raise normalized average precision by 10.1 points over prior best results.

Core claim

By using class-level text features as semantic anchors to identify aligned visual channels, suppress style-induced noise, and enlarge inter-class similarity margins, and by reformulating localization as a hierarchical autoregressive process that aligns deeper ViT layers with early coarse regression and shallower layers with later spatial refinement, prototype-based similarity learning overcomes class confusion and insufficient spatial cues to reach new state-of-the-art few-shot detection performance.

What carries the argument

Text-Anchored Semantic Mask (TSMa) selects class-intrinsic channels through visual-text channel interaction, while Stage-Aligned Hierarchical Autoregressive Regression (SHARe) progressively updates bounding boxes across ViT abstraction levels from deep to shallow.

If this is right

  • TSMa reduces class confusion by enlarging similarity margins between prototypes.
  • SHARe supplies the missing spatial information by using layer-specific cues at each regression stage.
  • The two components together enable training-free adaptation to novel classes with both higher classification and localization accuracy.
  • The gains are measured as a 10.1 point increase in normalized AP on the COCO few-shot benchmark.

Where Pith is reading between the lines

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

  • The same text-visual channel selection could be tested on other few-shot tasks such as instance segmentation where semantic alignment matters.
  • The autoregressive stage progression might be applied to standard detectors outside the few-shot regime to improve box quality.
  • If ViT layers naturally match task granularity, similar alignment could be explored for other vision pipelines that combine classification and localization.

Load-bearing premise

Class-level text features can reliably act as semantic anchors to select aligned visual channels and enlarge inter-class margins, and ViT layer-wise abstraction levels can be aligned with regression stages such that deeper layers guide coarse localization while shallower layers refine spatial details.

What would settle it

A controlled test in which removing the text-anchoring step leaves inter-class margins unchanged or in which swapping the ViT layer-to-stage assignment yields no drop in localization accuracy would falsify the claimed mechanisms.

Figures

Figures reproduced from arXiv: 2606.23069 by KunHo Heo, MyeongAh Cho, Seungjae Kim, Suyeon Kim, Wongyu Lee.

Figure 1
Figure 1. Figure 1: (a) Recent prototype-based similarity learning methods construct class proto￾types solely from visual features, and similarity scores computed with query features can cause inter-class margin collapse, leading to class confusion. (b) Our method con￾structs Semantic Prototypes by leveraging text features as anchors to retain only co￾activated channels, enlarging inter-class similarity margins and thereby im… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed framework. It consists of two main components: Text-Anchored Semantic Mask (TSMa), which constructs semantic prototypes by leveraging text embeddings as anchors, and Stage-Aligned Hierarchical Au￾toregressive Regression (SHARe), which injects multi-level ViT features in reverse order into each autoregressive regression stage for precise localization. 3.2 Text-Anchored Semantic Mask… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison with existing methods (30-shot). Qualitative Evaluation [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of ablation results on the proposed components. 4.4 Discussions [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Query–prototype similarity heatmaps of Top-5 class prototypes [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of average similarity scores. PiDiViT Ours 1: Zebra 2: Giraffe 3: Horse 4: Cow 5: Sheep Query (Zebra) Class Confusion [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of t-SNE visualization. Inter-Class Similarity Mar￾gin Enlargement via TSMa. To validate that TSMa re￾solves inter-class similarity mar￾gin collapse and improves class separability, we conduct anal￾yses on the COCO [23] test set. Tab. 6 reports Top-k Ac￾curacy, defined as the prob￾ability that the ground-truth class ranks among the top￾k most similar prototypes to the query feature, where our me… view at source ↗
Figure 8
Figure 8. Figure 8: Visual comparison of ViT feature injection strategies in SHARe [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
read the original abstract

Few-shot object detection aims to detect novel object categories from only a few labeled examples, avoiding costly large-scale annotation. Recent prototype-based similarity learning approaches enable training-free adaptation by matching query features with class prototypes. However, they suffer from two fundamental limitations: (i) class confusion arising from inter-class similarity margin collapse, and (ii) insufficient visual cues for precise localization, as similarity scores capture only class-level semantic affinity while providing limited spatial information. To address these issues, we introduce two complementary components. Text-Anchored Semantic Mask (TSMa) leverages class-level text features as semantic anchors to identify semantically aligned channels through channel-wise interaction between visual and text features. By suppressing style-induced spurious responses and emphasizing class-intrinsic signals, TSMa enlarges inter-class similarity margins and mitigates class confusion. We further propose Stage-Aligned Hierarchical Autoregressive Regression (SHARe), which reformulates localization as a hierarchical autoregressive process that progressively refines bounding boxes across multiple stages. SHARe leverages the layer-wise characteristics of ViT representations by aligning feature abstraction levels with regression stages: deeper layers guide early coarse localization, while shallower layers rich in edge and texture cues refine spatial details in later stages. Experiments on COCO demonstrate a new state of the art, outperforming the previous best by +10.1 nAP, with extensive analysis validating each component. The code is available at https://github.com/VisualScienceLab-KHU/ReSet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes Text-Anchored Semantic Mask (TSMa) and Stage-Aligned Hierarchical Autoregressive Regression (SHARe) to mitigate inter-class similarity margin collapse and insufficient spatial cues in prototype-based few-shot object detection. TSMa uses class-level text features (e.g., from CLIP) as semantic anchors for channel-wise interaction to suppress spurious responses and enlarge margins; SHARe reformulates localization as a hierarchical autoregressive process aligned with ViT layer abstraction levels. Experiments on COCO are reported to achieve new SOTA performance with a +10.1 nAP gain over prior best, supported by component analysis; code is released.

Significance. If the reported gains prove robust, the work would represent a meaningful advance in few-shot object detection by demonstrating effective integration of text-visual channel alignment and stage-wise regression with ViT features. The public code release strengthens reproducibility and enables direct follow-up.

major comments (3)
  1. [Abstract, §4] Abstract and §4 (experiments): The central +10.1 nAP claim is presented without tabulated baseline details, exact prior method name, shot setting (e.g., 1/5/10-shot), or error bars; this makes it impossible to verify whether the margin is driven by TSMa/SHARe or by unstated protocol choices.
  2. [§3.1] §3.1 (TSMa): The channel-selection mechanism assumes class-name text embeddings reliably identify semantically aligned visual channels for novel COCO categories; no ablation isolates performance under domain shift or when text-visual misalignment occurs, leaving the margin-enlargement claim load-bearing yet untested against the skeptic's concern.
  3. [§3.2] §3.2 (SHARe): The alignment of ViT layer abstraction levels with regression stages (deeper for coarse, shallower for refinement) is presented as a design choice without quantitative justification that this hierarchy outperforms standard multi-stage regression heads on the same backbone.
minor comments (2)
  1. [Abstract] Notation for nAP should be defined on first use (novel-class average precision).
  2. [§4] Figure captions for qualitative results should explicitly state the shot setting and whether TSMa/SHARe are ablated.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below with clarifications and proposed revisions to improve the manuscript.

read point-by-point responses

  1. Referee: [Abstract, §4] Abstract and §4 (experiments): The central +10.1 nAP claim is presented without tabulated baseline details, exact prior method name, shot setting (e.g., 1/5/10-shot), or error bars; this makes it impossible to verify whether the margin is driven by TSMa/SHARe or by unstated protocol choices.

    Authors: We agree that the presentation lacks sufficient detail for verification. In the revised manuscript, we will expand §4 with a comprehensive table listing the exact prior best method name, results broken down by 1-shot/5-shot/10-shot settings on COCO, and error bars from multiple runs. This will explicitly attribute the +10.1 nAP gain to the proposed components under standard protocols. revision: yes

  2. Referee: [§3.1] §3.1 (TSMa): The channel-selection mechanism assumes class-name text embeddings reliably identify semantically aligned visual channels for novel COCO categories; no ablation isolates performance under domain shift or when text-visual misalignment occurs, leaving the margin-enlargement claim load-bearing yet untested against the skeptic's concern.

    Authors: The COCO novel-class experiments already evaluate TSMa on categories outside typical CLIP training distributions, providing evidence of robustness. However, we acknowledge the absence of an explicit misalignment ablation. We will add a targeted discussion in §3.1 and a supplementary ablation using generic or mismatched text prompts to quantify sensitivity, while noting that full domain-shift experiments fall outside the current scope. revision: partial

  3. Referee: [§3.2] §3.2 (SHARe): The alignment of ViT layer abstraction levels with regression stages (deeper for coarse, shallower for refinement) is presented as a design choice without quantitative justification that this hierarchy outperforms standard multi-stage regression heads on the same backbone.

    Authors: We will add a new ablation study in §4 that directly compares SHARe against a standard multi-stage regression head applied to the identical ViT backbone features (without layer alignment). This will provide quantitative evidence supporting the hierarchical alignment choice. revision: yes

Circularity Check

0 steps flagged

No circularity: claims rest on external benchmark validation, not internal definitions or self-citations

full rationale

The paper introduces TSMa and SHARe as new modules for prototype-based few-shot detection, with performance asserted via COCO experiments (+10.1 nAP over prior best) rather than any derivation that reduces to fitted inputs or self-referential equations. No load-bearing steps invoke self-citations, uniqueness theorems from the same authors, or ansatzes smuggled via prior work; the text features and hierarchical regression are presented as design choices validated externally. The derivation chain is therefore self-contained against independent benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim depends on two domain assumptions about text features and ViT layer properties that are not derived or validated within the abstract; no free parameters or invented entities are explicitly introduced in the provided text.

axioms (2)
  • domain assumption Class-level text features from a pre-trained model can serve as reliable semantic anchors to identify aligned visual channels and suppress style-induced responses.
    Invoked to justify TSMa enlarging inter-class margins.
  • domain assumption ViT representations exhibit layer-wise characteristics in which deeper layers supply coarse localization cues and shallower layers supply edge and texture cues suitable for progressive refinement.
    Invoked to justify the stage alignment in SHARe.

pith-pipeline@v0.9.1-grok · 5805 in / 1483 out tokens · 53273 ms · 2026-06-30T10:35:27.215982+00:00 · methodology

Review history (2 revisions) →

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

44 extracted references · 6 canonical work pages · 3 internal anchors

  1. [1]

    ACM computing surveys (CSUR)54(11s), 1–37 (2022)

    Antonelli, S., Avola, D., Cinque, L., Crisostomi, D., Foresti, G.L., Galasso, F., Marini, M.R., Mecca, A., Pannone, D.: Few-shot object detection: A survey. ACM computing surveys (CSUR)54(11s), 1–37 (2022)

    2022

  2. [2]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Barsellotti,L.,Bianchi,L.,Messina,N.,Carrara,F.,Cornia,M.,Baraldi,L.,Falchi, F., Cucchiara, R.: Talking to dino: Bridging self-supervised vision backbones with language for open-vocabulary segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22025–22035 (2025)

    2025

  3. [3]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Bulat, A., Guerrero, R., Martinez, B., Tzimiropoulos, G.: Fs-detr: Few-shot detec- tion transformer with prompting and without re-training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11793–11802 (2023)

    2023

  4. [4]

    In:ProceedingsoftheIEEEconferenceoncomputervisionandpatternrecognition

    Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In:ProceedingsoftheIEEEconferenceoncomputervisionandpatternrecognition. pp. 1209–1218 (2018)

    2018

  5. [5]

    In: European conference on computer vision

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End- to-end object detection with transformers. In: European conference on computer vision. pp. 213–229. Springer (2020)

    2020

  6. [6]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

    2021

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, D.J., Hsieh, H.Y., Liu, T.L.: Adaptive image transformer for one-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12247–12256 (2021)

    2021

  8. [8]

    In: World Confer- ence on Explainable Artificial Intelligence

    Dorszewski, T., Tětková, L., Jenssen, R., Hansen, L.K., Wickstrøm, K.K.: From colors to classes: Emergence of concepts in vision transformers. In: World Confer- ence on Explainable Artificial Intelligence. pp. 28–47. Springer (2025)

    2025

  9. [9]

    Everingham,M.,VanGool,L.,Williams,C.K.,Winn,J.,Zisserman,A.:Thepascal visualobjectclasses(voc)challenge.Internationaljournalofcomputervision88(2), 303–338 (2010)

    2010

  10. [10]

    In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

    Fan, Q., Zhuo, W., Tang, C.K., Tai, Y.W.: Few-shot object detection with attention-rpn and multi-relation detector. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 4013–4022 (2020)

    2020

  11. [11]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Fan, Z., Ma, Y., Li, Z., Sun, J.: Generalized few-shot object detection without forgetting. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4527–4536 (2021)

    2021

  12. [12]

    In: European Conference on Computer Vision

    Fu, Y., Wang, Y., Pan, Y., Huai, L., Qiu, X., Shangguan, Z., Liu, T., Fu, Y., Van Gool, L., Jiang, X.: Cross-domain few-shot object detection via enhanced open-set object detector. In: European Conference on Computer Vision. pp. 247–

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Guirguis, K., Meier, J., Eskandar, G., Kayser, M., Yang, B., Beyerer, J.: Niff: Alleviating forgetting in generalized few-shot object detection via neural instance feature forging. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24193–24202 (2023) 16 K. Heoet al

    2023

  14. [14]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Han, G., He, Y., Huang, S., Ma, J., Chang, S.F.: Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3263–3272 (2021)

    2021

  15. [15]

    In: Proceedings of the AAAI conference on artificial intelligence

    Han, G., Huang, S., Ma, J., He, Y., Chang, S.F.: Meta faster r-cnn: Towards ac- curate few-shot object detection with attentive feature alignment. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 780–789 (2022)

    2022

  16. [16]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Han, G., Ma, J., Huang, S., Chen, L., Chang, S.F.: Few-shot object detection with fully cross-transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5321–5330 (2022)

    2022

  17. [17]

    In: Proceedings of the IEEE international conference on computer vision

    He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961–2969 (2017)

    2017

  18. [18]

    Advances in neural information processing systems32 (2019)

    Hsieh, T.I., Lo, Y.C., Chen, H.T., Liu, T.L.: One-shot object detection with co- attention and co-excitation. Advances in neural information processing systems32 (2019)

    2019

  19. [19]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot object detection via feature reweighting. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8420–8429 (2019)

    2019

  20. [20]

    Kaul, P., Xie, W., Zisserman, A.: Label, verify, correct: A simple few shot object detectionmethod.In:ProceedingsoftheIEEE/CVFconferenceoncomputervision and pattern recognition. pp. 14237–14247 (2022)

    2022

  21. [21]

    arXiv preprint arXiv:2509.17401 (2025)

    Kim,J.,Kim,J.,Shim,Y.,Kim,J.,Jung,S.,Hwang,S.J.:Interpretingvisiontrans- formers via residual replacement model. arXiv preprint arXiv:2509.17401 (2025)

    work page arXiv 2025

  22. [22]

    IEEE transactions on neural networks and learning systems35(9), 11958–11978 (2023)

    Köhler, M., Eisenbach, M., Gross, H.M.: Few-shot object detection: A compre- hensive survey. IEEE transactions on neural networks and learning systems35(9), 11958–11978 (2023)

    2023

  23. [23]

    In: European conference on computer vision

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740–755. Springer (2014)

    2014

  24. [24]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ma, J., Niu, Y., Xu, J., Huang, S., Han, G., Chang, S.F.: Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3208–3218 (2023)

    2023

  25. [25]

    Syngress Publishing, (2008)

    Manning, C.D.: Introduction to information retrieval. Syngress Publishing, (2008)

    2008

  26. [26]

    One-Shot Instance Segmentation

    Michaelis, C., Ustyuzhaninov, I., Bethge, M., Ecker, A.S.: One-shot instance seg- mentation. arXiv preprint arXiv:1811.11507 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  27. [27]

    Representation Learning with Contrastive Predictive Coding

    Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predic- tive coding. arXiv preprint arXiv:1807.03748 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  28. [28]

    DINOv2: Learning Robust Visual Features without Supervision

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  29. [29]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Qiao, L., Zhao, Y., Li, Z., Qiu, X., Wu, J., Zhang, C.: Defrcn: Decoupled faster r- cnn for few-shot object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8681–8690 (2021)

    2021

  30. [30]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  31. [31]

    Advances in neural information processing systems28(2015) ReSet 17

    Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object de- tection with region proposal networks. Advances in neural information processing systems28(2015) ReSet 17

    2015

  32. [32]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Sun, B., Li, B., Cai, S., Yuan, Y., Zhang, C.: Fsce: Few-shot object detection via contrastive proposal encoding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7352–7362 (2021)

    2021

  33. [33]

    arXiv preprint arXiv:2003.06957 (2020)

    Wang, X., Huang, T.E., Darrell, T., Gonzalez, J.E., Yu, F.: Frustratingly simple few-shot object detection. arXiv preprint arXiv:2003.06957 (2020)

    work page arXiv 2003

  34. [34]

    IEEE transactions on pattern analysis and machine intelligence45(3), 3090–3106 (2022)

    Xiao, Y., Lepetit, V., Marlet, R.: Few-shot object detection and viewpoint estima- tion for objects in the wild. IEEE transactions on pattern analysis and machine intelligence45(3), 3090–3106 (2022)

    2022

  35. [35]

    Information Fusion107, 102307 (2024)

    Xin, Z., Chen, S., Wu, T., Shao, Y., Ding, W., You, X.: Few-shot object detection: Research advances and challenges. Information Fusion107, 102307 (2024)

    2024

  36. [36]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Yan, X., Chen, Z., Xu, A., Wang, X., Liang, X., Lin, L.: Meta r-cnn: Towards gen- eral solver for instance-level low-shot learning. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9577–9586 (2019)

    2019

  37. [37]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yang, H., Cai, S., Sheng, H., Deng, B., Huang, J., Hua, X.S., Tang, Y., Zhang, Y.: Balanced and hierarchical relation learning for one-shot object detection. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7591–7600 (2022)

    2022

  38. [38]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45(12), 14832–14845 (2022)

    Zhang, G., Luo, Z., Cui, K., Lu, S., Xing, E.P.: Meta-detr: Image-level few-shot detection with inter-class correlation exploitation. IEEE Transactions on Pattern Analysis and Machine Intelligence45(12), 14832–14845 (2022)

    2022

  39. [39]

    arXiv preprint arXiv:2309.12969 (2023)

    Zhang, X., Liu, Y., Wang, Y., Boularias, A.: Detect everything with few examples. arXiv preprint arXiv:2309.12969 (2023)

    work page arXiv 2023

  40. [40]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhao, Y., Guo, X., Lu, Y.: Semantic-aligned fusion transformer for one-shot object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7601–7611 (2022)

    2022

  41. [41]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 16793–16803 (2022)

    2022

  42. [42]

    Uniform Average

    Zhou, H., Liu, Y., Mo, C., Li, W., Peng, B., Liu, L.: When pixel difference patterns meet vit: Pidivit for few-shot object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 24309–24318 (2025) Rethinking Prototype-based Similarity Learning for Few-Shot Object Detection Supplementary Material A Implementation Detail...

    2025

  43. [43]

    21. 22. 23. Input Image Input Image

  44. [44]

    21. 22. 23. Low - Level Features Mid - Level Features High - Level Features Fig.S2:Layer-wise similarity maps of ViT-L features. B.6 Layer-wise ViT Feature Visualization Fig. S2 visualizes the cosine similarity between the CLS token and patch tokens at each of the 24 layers of ViT-L, where the CLS token serves as a global rep- resentation of the image. Th...