Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Perceptual Anchoring: Prototype-Guided Text Calibration for Training-free Open-Vocabulary Semantic Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that calibrating generic class text embeddings with per-image visual prototypes—built from the tokens with the highest first-pass score margins—reduces the semantic gap in training-free open-vocabulary semantic…

desk verdict A simple, clearly motivated text-calibration idea with real but small gains, whose plug-and-play claim is undercut by per-dataset tuning of its main hyperparameters. read the letter →

arxiv 2608.03991 v2 pith:HJTA4JFQ submitted 2026-08-04 cs.CV

Wanli Ma , Jiangwen Lu , Qinmu Peng , Xinge You This is my paper · ORCID
classification cs.CV
keywords open-vocabularysemanticsegmentationtraining-freeCLIPtextcalibrationvisualprototypeperceptualanchoringgapplug-and-play
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Training-free open-vocabulary segmentation matches each image patch against text embeddings that describe a category in the abstract, and this paper argues that the mismatch between those generic embeddings and the image's specific appearances is a real cause of incomplete and wrong masks. The proposed fix, Prototype-Guided Text Calibration (PTC), works in two stages: it first selects the most reliable visual tokens per predicted class using the margin between the top two matching scores, and then blends the resulting category prototype into the class text embedding with a strength that grows with the amount of supporting evidence. This is a plug-and-play module: no training, no extra models, and it can sit on top of existing training-free segmenters. The paper reports that adding PTC to six baselines improves average mIoU by 0.8 to 2.2 points across eight benchmarks, with the largest gains on datasets where objects are salient and categories are few. If the claim is right, per-image text calibration is a generally useful complement to the visual-feature improvements that most prior work focuses on.

What carries the argument

The load-bearing identity is the evidence-driven convex combination in Eq. (11): $t_{\mathrm{cal}}^c = (1-\mu_c)\,t_c + \mu_c\,V_{\mathrm{proto}}^c$. The prototype $V_{\mathrm{proto}}^c$ is a margin-weighted average of the visual tokens whose initial matching scores put them in class $c$ with the largest margin between the top and second class, so it encodes what that specific instance looks like. The coefficient $\mu_c = \mu\,\min\{1,\log(1+n_{\mathrm{ev}}^c)/\log(1+\lambda K_{\min})\}$ keeps the calibration conservative when evidence is thin and prevents large objects from pulling the text embedding too far from its generic meaning. This single interpolation rule carries the argument from the Perceiving stage to the Anchoring stage.

What would settle it

Choose images in which a training-free segmenter gives a large, confident margin to background or to a confusable region for a class (for instance a black dog against a dark wall), build the PTC prototype from those high-margin tokens, and check whether mIoU falls below the uncalibrated baseline. If PTC systematically amplifies confident baseline errors instead of correcting them, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the semantic gap in training-free OVSS can be narrowed by treating text embeddings as anchor symbols that should be perceptually linked to the input image, rather than as fixed classification references. Its concrete proposal is the calibration rule $t_{\mathrm{cal}}^c = (1-\mu_c)\,t_c + \mu_c\,V_{\mathrm{proto}}^c$ for every class with enough evidence, where $V_{\mathrm{proto}}^c$ is a visual prototype formed by aggregating the $K_c$ visual tokens with the largest score margins for that class, weighted by margin size. Categories with fewer than $K_{\min}$ supporting tokens are left uncalibrated, and the calibration strength $\mu_c$ grows logarithmically with evidence until it saturates at a global bound $\mu$. The paper validates this rule by integrating it into six training-free baselines and reporting mIoU gains across the eight benchmarks, and by ablations showing that margin-based selection, the hybrid evidence-quantity rule, and adaptive strength each contribute to the gain.

Load-bearing premise

The method assumes that the image regions its starting model is most confident about really do belong to the predicted class, so if that model is confidently wrong, the calibration will build on and reinforce the same mistake.

Editorial extensions

If this is right

  • PTC can be added to any training-free OVSS method that emits per-token matching scores, so future visual-feature improvements and this text-side calibration should stack rather than compete.
  • Because the calibrated embedding remains a blend with the original generic text vector, the open-vocabulary behavior of the model is preserved for categories that are absent from the image.
  • The gain is largest where the baseline is weakest (NACLIP +2.2 average mIoU) and smallest on the strongest baseline (CorrCLIP +0.9), suggesting that text calibration and strong visual representations address overlapping but not identical errors.
  • The same recipe should transfer across vision foundation models and CLIP backbones, since the ablations report gains with SAM, DINO, and DINOv2 features and with ViT-B/16, ViT-L/14, and ViT-H/14 encoders.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not explore iterative recalibration, so an untested consequence of the margin-based evidence filter is that using calibrated embeddings to produce new margins could re-anchor prototypes and further close the gap.
  • A transfer the paper leaves implicit is to other CLIP-based dense prediction tasks, such as open-vocabulary detection or referring segmentation, where generic text embeddings are matched to region features.
  • The dataset dependence of $K_{\min}$ and $\mu$ suggests that a self-tuning schedule for these two hyperparameters would be the natural next step toward a fully parameter-free plug-in.
  • A stress test implied by the paper's own limitation statement is to feed PTC deliberately biased initial scores, to see whether the margin filter or the calibration rule is the main source of robustness.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Prototype-Guided Text Calibration (PTC), a training-free module for open-vocabulary semantic segmentation that calibrates category text embeddings with image-specific visual prototypes. In the Perceiving stage, PTC selects high-margin visual tokens as evidence and aggregates them into per-category prototypes; in the Anchoring stage, it blends these prototypes into the text embeddings with an evidence-adaptive strength, leaving categories with insufficient evidence unchanged. The method is evaluated by adding PTC to six training-free baselines across eight benchmarks, reporting average mIoU gains of roughly 0.8 to 2.2 points, together with ablations of the reliability criterion, the evidence-quantity strategy, the calibration scheme, and the choice of foundation backbones.

Significance. If the reported gains are robust, PTC is a simple and attractive contribution: it is algorithmically clean, requires no training and no additional external models, and the ablations in Tables 2-4 isolate each design choice. The authors also ship code and evaluate across six baselines and eight datasets, which is a strength. However, the central empirical claim is currently weaker than the text suggests because the headline results appear to use dataset-specific and baseline-specific hyperparameters, and because several reported gains are within the expected run-to-run variance. The idea itself is sensible and the derivation is not circular, but the paper must demonstrate a fixed-configuration, plug-and-play version before the generalization claim at the claimed strength can be accepted.

major comments (4)
  1. [§4.3.2, Fig. 5, Table 1] The sensitivity analysis in §4.3.2 shows that the optimal Kmin and μ vary by orders of magnitude across datasets (Kmin=100 on VOC21, Kmin=2 on Cityscapes, Kmin=10 on ADE20K, Kmin=132 on COCO-Object; μ=0.30, 0.35, 0.15, 0.02 respectively), and the text says 'we adopt dataset-specific configurations for the current baseline' and 'these two parameters are adjusted when PTC is integrated with other baselines.' This means the Table 1 entries likely use per-dataset, per-baseline tuned hyperparameters, so the average gains of +0.8 to +2.2 mIoU are an upper bound rather than evidence that PTC is a plug-and-play module. Please add an evaluation with one fixed (Kmin, μ) configuration held constant across all datasets and baselines, or a nested cross-validation procedure on the validation splits, and report the resulting average gains.
  2. [Table 1 and §4.3] No repeated-run variance is reported, and several individual improvements are tiny: ResCLIP+PTC gains +0.0 on Context60, ClearCLIP+PTC gains +0.1 on VOC21 and COCO-Object, and CorrCLIP+PTC gains +0.1 on Cityscapes and +0.0 on ADE20K. Single deterministic runs cannot distinguish these from run-to-run noise or numerical nondeterminism. Please report means and standard deviations over at least three runs, or otherwise provide evidence that the small gains are statistically meaningful; currently the average improvements inherit these small, possibly non-significant increments.
  3. [§5, §4.3.3, Table 6] The conclusion acknowledges that 'the performance of PTC remains dependent on the quality of the constructed visual prototypes.' This is not merely a caveat: Table 6 already shows a degradation on COCO-Object when ViT-L/14 is used (39.2 to 38.8). Since the prototypes are built from the very initial predictions that PTC is trying to correct, biased or background-dominated initial scores can propagate into the calibrated text embeddings. Please add a failure analysis or an explicit condition under which PTC is expected to help versus hurt, such as a per-category breakdown or a comparison of prototypes constructed from correctly versus incorrectly classified initial tokens.
  4. [§4.3.3, Table 5] The text states that 'incorporating PTC consistently improves segmentation performance across all three VFM variants,' but Table 5 shows SAM-B/16 +PTC decreasing ADE20K from 19.1 to 18.9. This overstatement should be corrected, and the failure mode should be discussed, because the robustness claim across visual foundation models is part of the paper's generalization argument.
minor comments (5)
  1. [§3.2.3, Eq. (11)] The second case of Eq. (11) is written 'tc,c<Cvalid', which should be 'tc, c∉Cvalid' to denote categories not in the valid set.
  2. [Title] The title contains a typo, 'Open-V ocabulary', which should be 'Open-Vocabulary'.
  3. [Fig. 5] The text in Figure 5, especially the Kmin x-axis labels, is very small and difficult to read at print size; please increase font sizes and consider annotating the best operating points directly in the figure.
  4. [§4.1] The implementation extracts visual evidence from overlapping 224x224 local crops and aggregates it across crops, but Section 3.2 formulates the method on a single set of image features and does not specify how per-crop tokens are aggregated or how overlapping regions are handled; please reconcile the description with the implementation or provide pseudocode for the crop aggregation step.
  5. [Table 1] Several entries in Table 1 are marked with dashes; please clarify in the caption or text whether dashes mean the method was not evaluated on that dataset, the result is not reported, or the dataset is not applicable to the method.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: PTC is self-referential as test-time adaptation, but the reported gains are empirical outcomes, not consequences of the defining equations.

full rationale

The paper's derivation chain is not circular in the sense targeted by this review. PTC constructs visual prototypes from initial matching scores (Eqs. 4-8) and defines calibrated text embeddings as a convex combination of generic text embeddings and these prototypes (Eq. 11). This is self-referential in that the image's own predictions are recycled to adjust the classifier, but the final segmentation maps are computed from the calibrated embeddings and evaluated against external ground-truth labels. The claimed mIoU improvements are therefore empirical results, not consequences of the definitions. Eq. (11) does not reduce Eq. (3) to Eq. (1) by construction, and no fitted parameter is renamed as a prediction: Kmin, mu, lambda, and rho are manually specified hyperparameters whose effects are analyzed in Sec. 4.3, and the benchmark numbers are measured on the standard validation sets. The dataset-specific hyperparameter choices in Sec. 4.3.2 may weaken the 'plug-and-play' generality claim, but that is an evaluation-robustness concern, not an input-output equivalence. The stated limitation in Sec. 5 ('the performance of PTC remains dependent on the quality of the constructed visual prototypes') is a real dependency, but it describes the method's operating condition rather than a circular derivation. No load-bearing self-citations or imported uniqueness theorems appear in the paper.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. It relies on four domain assumptions about CLIP features and one set of benchmark-tuned hyperparameters. The most fragile assumption is that margin-ranked tokens are reliable evidence, which the paper itself flags in the conclusion.

free parameters (4)
  • Kmin (minimum evidence tokens) = 100 (VOC21), 2 (Cityscapes), 10 (ADE20K), 132 (COCO-Object); adjusted per baseline
    Controls the minimum visual evidence for a category to be calibrated; tuned per dataset and per baseline on validation sets.
  • mu (global calibration strength) = 0.3 (VOC21), 0.35 (Cityscapes), 0.15 (ADE20K), 0.02 (COCO-Object); adjusted per baseline
    Global upper bound on calibration strength; tuned per dataset and per baseline; large effect on scores.
  • lambda (evidence saturation factor) = 10
    Selected by ablation (5 vs 10 vs 15) on the same six datasets; controls how many evidence tokens are needed for calibration strength to saturate.
  • rho (sampling ratio) = 0.1
    Sampling ratio used in the hybrid evidence quantity; chosen from variants 0.1, 0.2, 0.3 by validation results.
assumptions (4)
  • domain assumption CLIP text and image embeddings are commensurable via cosine similarity.
    The whole pipeline (Eqs. 1-3) relies on cosine matching between text and visual features.
  • domain assumption Tokens with the largest score margin are reliable visual evidence for their predicted class.
    Margin-based selection is the core of Perceiving (Eqs. 5-7); if false, prototypes inherit baseline errors.
  • domain assumption The convex combination in Eq. (11) preserves generic category semantics while adding instance detail.
    No proof is given that the interpolated embedding remains a valid text reference; it is validated only indirectly by mIoU.
  • domain assumption CLIP patch features retain object-level localization after the modifications of the baseline.
    Inherited from the CLIP-based dense prediction literature and used to build prototypes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Perceptual Anchoring: Prototype-Guided Text Calibration for Training-free Open-Vocabulary Semantic Segmentation." pith.science (2026). https://pith.science/paper/HJTA4JFQ

@misc{pith2026260803991,
  author       = {Pith},
  title        = {Pith review of: Perceptual Anchoring: Prototype-Guided Text Calibration for Training-free Open-Vocabulary Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HJTA4JFQ}},
  note         = {Machine review of arXiv:2608.03991}
}
read the original abstract

Training-free open-vocabulary semantic segmentation (OVSS) partitions an image into semantically distinct regions based on arbitrary text descriptions, without learning any additional parameters. However, existing methods typically focus on improving visual representations while treating text embeddings that encode only generic category concepts as fixed classification references. The resulting semantic gap between these generic concepts and the visual representations that capture the specific appearances of target instances often causes incomplete masks and erroneous predictions in non-target regions. Inspired by the symbol-percept correspondence underlying perceptual anchoring, we propose Prototype-Guided Text Calibration (PTC) for training-free OVSS. In the Perceiving stage, PTC selects reliable visual evidence based on initial matching scores to construct category-specific visual prototypes. In the Anchoring stage, PTC uses these prototypes to calibrate their corresponding text embeddings, with the calibration strength adaptively adjusted based on the amount of visual evidence. Consequently, the calibrated text embeddings align more accurately with instance-specific visual representations while preserving generic category semantics and open-vocabulary generalization. Moreover, PTC requires neither additional training nor external models and can serve as a plug-and-play module for existing methods. Extensive experiments across eight benchmarks show that PTC significantly enhances the performance of six representative methods and yields more complete and accurate segmentation results. These results validate PTC as a simple and effective approach to improving visual-text alignment.

Figures

Figures reproduced from arXiv: 2608.03991 by the authors.

Figure 1
Figure 1. Motivation and effectiveness of PTC. (a) Compared with the ProxyCLIP baseline, adding PTC yields more complete activation over object regions and suppresses background-biased responses. (b) Introducing PTC consistently improves ProxyCLIP across all eight benchmarks. [18, 11, 19, 20, 21, 22, 23, 24, 12, 25, 26, 27, 28]. However, these methods often treat class text embeddings as reliable zero-shot classifiers and dir… view at source ↗
Figure 2
Figure 2. Overall framework of PTC for training-free OVSS. 1. Feature Extraction: The input image and class names are encoded by frozen encoders to obtain visual features and text embeddings, respectively. 2. PTC: In the Perceiving stage, reliable visual evidence is selected from the visual features based on initial matching scores and aggregated into visual prototypes. In the Anchoring stage, the text embeddings of categorie… view at source ↗
Figure 3
Figure 3. Visualization results on the PASCAL VOC dataset. “GT” denotes ground truth. Image ProxyCLIP + PTC GT [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization results on the COCO-Object dataset. “GT” denotes ground truth. these interferences, producing more accurate predictions. In Example 9, PTC successfully captures small objects that are missed by ProxyCLIP. Moreover, PTC improves the segmentation of a black…
Figure 5
Figure 5. Figure 5: Sensitivity Analysis of Kmin and µ. Top row: Effect of the minimum quantity of visual evidence (Kmin) on mIoU across four datasets. Bottom row: Effect of the global calibration strength (µ). 4.3.2. Sensitivity Analysis of Kmin and µ In this subsection, we analyze the e…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 3 canonical work pages

  1. [1]

    L.-C. Chen, G. Papandreou, I. Kokkinos, et al., Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs, IEEE Trans. Pattern Anal. Mach. Intell. 40 (2018) 834–848. doi:10.1109/TPAMI.2017.2699184

    arXiv 2018

  2. [2]

    E. Xie, W. Wang, Z. Yu, et al., Segformer: Simple and efficient design for semantic segmentation with transformers, in: NeurIPS, 2021, V ol. 34, 2021, pp. 12077–12090.doi:10.5555/3540261.3541185

    arXiv 2021

  3. [4]

    Guo, C.-Z

    M.-H. Guo, C.-Z. Lu, Q. Hou, et al., Segnext: Rethinking convolutional attention design for semantic segmentation, in: NeurIPS, 2022, V ol. 35, pp. 1140–1156.doi:10.5555/3600270.3600354

    arXiv 2022

  4. [5]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, et al., Learning transferable visual models from natural language supervision, in: ICML, 2021, V ol. 139, pp. 8748–8763

    2021

  5. [6]

    C. Jia, Y . Yang, Y . Xia, et al., Scaling up visual and vision-language representation learning with noisy text supervision, in: ICML, 2021, V ol. 139, pp. 4904–4916. 14

    work page 2021

  6. [8]

    S. Jiao, Y . Wei, Y . Wang, et al., Learning mask-aware clip representations for zero-shot segmentation, in: NeurIPS, 2023, V ol. 36, pp. 35631–35653.doi:10.52202/075280-1549

    doi:10.52202/075280-1549 2023

  7. [9]

    J. Xu, S. D. Mello, S. Liu, et al., Groupvit: Semantic segmentation emerges from text supervision, in: CVPR, 2022, pp. 18113–18123.doi:10.1109/CVPR52688.2022.01760

    work page arXiv 2022

  8. [10]

    Y . Xing, J. Kang, A. Xiao, et al., Rewrite caption semantics: Bridging semantic gaps for language-supervised semantic segmentation, in: NeurIPS, 2023, V ol. 36, pp. 68798–68809.doi:10.52202/075280-3011

    doi:10.52202/075280-3011 2023

Show all 45 references
  1. [11]

    F. Wang, J. Mei, A. L. Yuille, SCLIP: rethinking self-attention for dense vision-language inference, in: ECCV, 2024, V ol. 15079, pp. 315–332.doi:10.1007/978-3-031-72664-4_18

    2024 doi

  2. [12]

    M. Lan, C. Chen, Y . Ke, et al., Proxyclip: Proxy attention improves CLIP for open-vocabulary segmentation, in: ECCV, 2024, V ol. 15126, pp. 70–88.doi:10.1007/978-3-031-73113-6_5

    2024 doi

  3. [13]

    S. Jiao, H. Zhu, J. Huang, et al., Collaborative vision-text representation optimizing for open-vocabulary segmen- tation, in: ECCV, 2024, V ol. 15091, pp. 399–416.doi:10.1007/978-3-031-73414-4_23

    2024 doi

  4. [17]

    S. Wu, W. Zhang, L. Xu, et al., Clipself: Vision transformer distills itself for open-vocabulary dense prediction, in: B. Kim, Y . Yue, S. Chaudhuri, et al. (Eds.), ICLR 2024, V ol. 2024, pp. 52484–52503.doi:10.48550/arXiv. 2310.01403

    2024 doi arXiv

  5. [18]

    C. Zhou, C. C. Loy, B. Dai, Extract free dense labels from CLIP, in: ECCV, 2022, V ol. 13688, pp. 696–712. doi:10.1007/978-3-031-19815-1_40

    2022 doi

  6. [19]

    Hajimiri, I

    S. Hajimiri, I. B. Ayed, J. Dolz, Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation, in: W ACV, 2025, 2025, pp. 5061–5071.doi:10.1109/WACV61041.2025.00495

    2025

  7. [20]

    Y . Li, H. Wang, Y . Duan, et al., A closer look at the explainability of contrastive language-image pre-training, Pattern Recognition 162 (2025) 111409.doi:10.1016/j.patcog.2025.111409

    2025

  8. [22]

    T. Shao, Z. Tian, H. Zhao, et al., Explore the potential of clip for training-free open vocabulary semantic segmentation, in: ECCV, 2024, V ol. 15144, pp. 139–156.doi:10.1007/978-3-031-73016-0_9

    2024 doi

  9. [23]

    M. Lan, C. Chen, Y . Ke, et al., Clearclip: Decomposing CLIP representations for dense vision-language inference, in: ECCV, 2024, V ol. 15105, pp. 143–160.doi:10.1007/978-3-031-72970-6_9

    2024 doi

  10. [24]

    Y . Yang, J. Deng, W. Li, et al., Resclip: Residual attention for training-free dense vision-language inference, in: CVPR, 2025, pp. 29968–29978.doi:10.1109/CVPR52734.2025.02789

    2025

  11. [27]

    Caron, H

    M. Caron, H. Touvron, I. Misra, et al., Emerging properties in self-supervised vision transformers, in: ICCV, 2021, pp. 9650–9660.doi:10.1109/ICCV48922.2021.00951

    2021

  12. [29]

    Radford, J

    A. Radford, J. Wu, R. Child, et al., Language models are unsupervised multitask learners (2019)

    2019

  13. [30]

    J. Lee, K. Toutanova, Pre-training of deep bidirectional transformers for language understanding, in: NAACL, 2019, V ol. 1, pp. 4171–4186.doi:10.18653/V1/N19-1423

    2019 doi

  14. [31]

    Coradeschi, A

    S. Coradeschi, A. Saffiotti, Perceptual Anchoring of Symbols for Action, in: IJCAI, 2001, pp. 407–416

    2001

  15. [32]

    Coradeschi, A

    S. Coradeschi, A. Saffiotti, An introduction to the anchoring problem, Robotics and autonomous systems 43 (2-3) (2003) 85–96

    2003

  16. [33]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, et al., Dinov2: Learning robust visual features without supervision, arXiv preprint arXiv:2304.07193 (2023)

    2023 arXiv

  17. [35]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, et al., High-resolution image synthesis with latent diffusion models, in: CVPR, 2022, pp. 10684–10695.doi:10.48550/ARXIV.2204.11824

    2022 doi arXiv

  18. [36]

    Barsellotti, R

    L. Barsellotti, R. Amoroso, L. Baraldi, et al., Fossil: Free open-vocabulary semantic segmentation through synthetic references retrieval, in: W ACV, 2024, pp. 1464–1473.doi:10.1109/WACV57701.2024.00149

    2024

  19. [38]

    W. Sun, Y . Du, G. Liu, et al., Training-free semantic segmentation via llm-supervision, arXiv preprint arXiv:2404.00701 (2024)

    2024 arXiv

  20. [39]

    Brown, B

    T. Brown, B. Mann, N. Ryder, et al., Language models are few-shot learners, in: NeurIPS, 2020, V ol. 33, pp. 1877–1901.doi:10.5555/3495724.3495883

    2020

  21. [40]

    M. A. Aydın, E. M. Çırpar, E. Abdinli, et al., Itaclip: Boosting training-free semantic segmentation with image, text, and architectural enhancements, in: CVPR Workshops, 2025, pp. 4181–4191. doi:10.1109/CVPRW67362. 2025.00398

    2025

  22. [41]

    Q. Chen, L. Yang, Y . Chen, et al., Training-free class purification for open-vocabulary semantic segmentation, in: ICCV, 2025, pp. 23124–23134.doi:10.1109/ICCV51701.2025.02147

    2025

  23. [42]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, et al., Attention is all you need, in: NeurIPS, 2017, V ol. 30, 2017, pp. 6000 – 6010.doi:10.5555/3295222.3295349

    2017

  24. [43]

    Everingham, S

    M. Everingham, S. M. A. Eslami, L. V . Gool, et al., The pascal visual object classes challenge: A retrospective, Int. J. Comput. Vis. 111 (1) (2015) 98–136.doi:10.1007/S11263-014-0733-5

    2015 doi

  25. [44]

    Mottaghi, X

    R. Mottaghi, X. Chen, X. Liu, et al., The role of context for object detection and semantic segmentation in the wild, in: CVPR, 2014, pp. 891–898.doi:10.1109/CVPR.2014.119

    2014 doi

  26. [45]

    Caesar, J

    H. Caesar, J. Uijlings, V . Ferrari, Coco-stuff: Thing and stuffclasses in context, in: CVPR, 2018, pp. 1209–1218. doi:10.1109/CVPR.2018.00132. 16

    2018

  27. [46]

    Cordts, M

    M. Cordts, M. Omran, S. Ramos, et al., The cityscapes dataset for semantic urban scene understanding, in: CVPR, 2016, pp. 3213–3223.doi:10.1109/CVPR.2016.350

    2016 doi

  28. [47]

    B. Zhou, H. Zhao, X. Puig, et al., Semantic understanding of scenes through the ADE20K dataset, Int. J. Comput. Vis. 127 (3) (2019) 302–321.doi:10.1007/S11263-018-1140-0

    2019 doi

  29. [48]

    G. Shin, W. Xie, S. Albanie, Reco: Retrieve and co-segment for zero-shot transfer, in: NeurIPS, 2022, V ol. 35, pp. 33754–33767.doi:10.52202/068431-2446

    2022 doi

  30. [49]

    S. Sun, R. Li, P. Torr, et al., Clip as rnn: Segment countless visual concepts without training endeavor, in: CVPR, 2024, pp. 13171–13182.doi:10.1109/CVPR52733.2024.01251

    2024

  31. [50]

    Karazija, I

    L. Karazija, I. Laina, A. Vedaldi, et al., Diffusion models for open-vocabulary segmentation, in: ECCV, 2024, V ol. 15099, pp. 299–317.doi:10.1007/978-3-031-72652-1_18

    2024 doi

  32. [51]

    Rewatbowornwong, N

    P. Rewatbowornwong, N. Chatthee, E. Chuangsuwanich, et al., Zero-guidance segmentation using zero segment labels, in: ICCV, 2023, pp. 1162–1172.doi:10.1109/ICCV51070.2023.00113

    2023

  33. [52]

    D. Kang, M. Cho, In defense of lazy visual grounding for open-vocabulary semantic segmentation, in: ECCV, 2024, V ol. 15099, pp. 143–164.doi:10.1007/978-3-031-72940-9_9

    2024 doi

  34. [53]

    Contributors, MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark, https:// github.com/open-mmlab/mmsegmentation(2020)

    M. Contributors, MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark, https:// github.com/open-mmlab/mmsegmentation(2020)

    2020

  35. [54]

    Contributors, MMEngine: Openmmlab foundational library for training deep learning models, https: //github.com/open-mmlab/mmengine(2022)

    M. Contributors, MMEngine: Openmmlab foundational library for training deep learning models, https: //github.com/open-mmlab/mmengine(2022)

    2022

  36. [55]

    Krähenbühl, V

    P. Krähenbühl, V . Koltun, Efficient inference in fully connected crfs with gaussian edge potentials, in: NeurIPS, 2011, pp. 109–117.doi:10.5555/2986459.2986472

    2011

  37. [56]

    Araslanov, S

    N. Araslanov, S. Roth, Single-stage semantic segmentation from image labels, in: CVPR, 2020, pp. 4252–4261. doi:10.1109/CVPR42600.2020.00431. 17

    2020

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.