pith. sign in

arxiv: 2606.09474 · v1 · pith:WPB4ECLFnew · submitted 2026-06-08 · 💻 cs.CV

Training-Free Generalized Few-Shot Segmentation through Open-Vocabulary Semantic Arbitration

Silas Kwabla Gah , Ebenezer Owusu This is my paper

Pith reviewed 2026-06-27 17:10 UTC · model grok-4.3

classification 💻 cs.CV
keywords generalized few-shot segmentationopen-vocabulary segmentationtraining-freeSAMCLIPsemantic arbitrationfoundation models
0
0 comments X

The pith

Open-V achieves 77.9 harmonic mIoU on PASCAL-5i for generalized few-shot segmentation without any training by arbitrating frozen SAM and CLIP priors.

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

The paper asks whether generalized few-shot semantic segmentation must rely on task-specific training or can instead be solved by coordinating existing foundation models at inference time. It introduces Open-V, which combines SAM3 Promptable Concept Segmentation with a K-shot CLIP support centroid using a calibrated per-pixel semantic arbitration rule. On the PASCAL-5i 1-shot benchmark this yields base/novel/harmonic mIoU scores of 78.4/77.5/77.9 and exceeds the strongest trained baseline by 17.7 points in harmonic mean. The work further shows that support examples contribute more when text priors are weak on disjoint vocabularies and that standard preprocessing mismatches can inflate prior results. The same training-free coordination generalizes to COCO-20i and ADE-OW under both conventional and open-vocabulary protocols.

Core claim

Open-V is a training-free GFSS framework that combines Segment Anything (SAM3) Promptable Concept Segmentation with a K-shot CLIP support centroid through calibrated per-pixel semantic arbitration, attaining base/novel/harmonic mIoU of 78.4/77.5/77.9 on PASCAL-5i (1-shot) without GFSS-specific training and surpassing the strongest trained baseline by +17.7 HM.

What carries the argument

Calibrated per-pixel semantic arbitration that grounds SAM3 promptable concept segmentation outputs against a K-shot CLIP support centroid derived from limited examples.

Load-bearing premise

The open-vocabulary priors inside frozen SAM and CLIP are sufficiently aligned and complementary that a simple calibrated per-pixel arbitration rule can accurately segment novel classes without any parameter updates.

What would settle it

Replace the calibrated arbitration rule with a naive average or threshold of the same SAM and CLIP outputs on PASCAL-5i 1-shot and measure whether harmonic mIoU falls below the strongest trained baseline.

Figures

Figures reproduced from arXiv: 2606.09474 by Ebenezer Owusu, Silas Kwabla Gah.

Figure 1
Figure 1. Figure 1: Open-Vpipeline. Frozen foundation models (SAM3 ViT-L, CLIP ViT-B/16) carry both evidence sources; the K-shot path on the right is applied only to novel classes; the per-pixel arg-max at the bottom is the arbitration stage. The query is en￾coded once by SAM3; SAM3-PCS decodes one text-conditioned instance set per class c ∈ Cb∪Cn. Base-class scores pass through unchanged; novel-class scores are reranked agai… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative strict-GFSS predictions on PASCAL- [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
read the original abstract

Generalized Few-Shot Semantic Segmentation (GFSS) has traditionally been approached as a representation-learning problem, requiring task-specific adaptation to incorporate novel classes from limited support examples. Recent foundation models, however, already exhibit strong open-vocabulary recognition and segmentation capabilities, raising a different question: can GFSS be solved through inference-time coordination of frozen semantic priors rather than parameter adaptation? We answer this question with Open-V, a training-free GFSS framework that combines Segment Anything (SAM3) Promptable Concept Segmentation (PCS) with a K-shot CLIP support centroid through calibrated per-pixel semantic arbitration. OpenV introduces no trainable components and supports arbitrary semantic categories at inference time. Beyond segmentation performance, our study contributes three broader findings. First, we show that support information can be incorporated through inference-time semantic grounding, and that its contribution increases as foundation-model text priors weaken on label-disjoint vocabularies. Second, we identify a reproducibility confound in foundationmodel segmentation, demonstrating that preprocessing and evaluation-space mismatches can silently distort reported performance. Finally, we validate Open-V across PASCAL5i, COCO-20i, and ADE-OW, showing that training-free coordination of foundation-model priors generalizes across both conventional GFSS and open-vocabulary evaluation settings. On PASCAL-5i (1-shot), Open-V attains base/novel/harmonic mIoU of 78.4/77.5/77.9, without GFSS-specific training surpassing the strongest trained baseline by +17.7 HM.

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

2 major / 2 minor

Summary. The manuscript presents Open-V, a training-free GFSS framework that fuses SAM3 Promptable Concept Segmentation with K-shot CLIP support centroids via calibrated per-pixel semantic arbitration of frozen open-vocabulary priors. It reports base/novel/harmonic mIoU of 78.4/77.5/77.9 on PASCAL-5i (1-shot), surpassing the strongest trained baseline by +17.7 HM, and validates the approach on COCO-20i and ADE-OW while identifying preprocessing confounds in foundation-model evaluations.

Significance. If the results hold under a fully specified, non-tuned arbitration rule, the work would be significant for demonstrating that inference-time coordination of existing foundation-model priors can solve GFSS without representation learning or task-specific optimization. The reproducibility analysis on preprocessing mismatches is a useful secondary contribution.

major comments (2)
  1. [Abstract/Method] Abstract and Method: the central claim that Open-V is training-free rests on 'calibrated per-pixel semantic arbitration' fusing SAM3 PCS masks with the CLIP centroid, yet no equation, decision rule, or pseudocode is supplied for the arbitration function or the calibration procedure. This is load-bearing; without it, one cannot verify whether the calibration is analytic/fixed or contains implicit dataset-level choices that would violate the no-optimization guarantee.
  2. [§4] §4 Experiments: the reported +17.7 HM gain and the three broader findings are presented without ablations on the arbitration calibration parameter, without error bars, and without verification that the gains survive identical preprocessing pipelines for all baselines. These omissions prevent assessment of whether the performance advantage is robust or confounded.
minor comments (2)
  1. [Abstract] The abstract states that three broader findings are contributed but does not enumerate them; listing them explicitly would improve clarity.
  2. [Method] Notation for the support centroid and per-pixel arbitration scores should be introduced with explicit symbols rather than descriptive phrases only.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback emphasizing reproducibility and full specification of the arbitration procedure. We address the major comments point by point below.

read point-by-point responses

  1. Referee: [Abstract/Method] Abstract and Method: the central claim that Open-V is training-free rests on 'calibrated per-pixel semantic arbitration' fusing SAM3 PCS masks with the CLIP centroid, yet no equation, decision rule, or pseudocode is supplied for the arbitration function or the calibration procedure. This is load-bearing; without it, one cannot verify whether the calibration is analytic/fixed or contains implicit dataset-level choices that would violate the no-optimization guarantee.

    Authors: We agree that the arbitration function requires explicit specification to substantiate the training-free claim. The procedure is a fixed analytic rule (per-pixel max-similarity comparison between SAM3 PCS logits and the K-shot CLIP centroid, with a constant threshold derived solely from support-set statistics and no dataset-level optimization). We will add the full equation, decision rule, and pseudocode to the Method section in revision. revision: yes

  2. Referee: [§4] §4 Experiments: the reported +17.7 HM gain and the three broader findings are presented without ablations on the arbitration calibration parameter, without error bars, and without verification that the gains survive identical preprocessing pipelines for all baselines. These omissions prevent assessment of whether the performance advantage is robust or confounded.

    Authors: The comment is correct that ablations, error bars, and explicit preprocessing verification are absent. We will add (i) an ablation varying the single calibration threshold, (ii) standard error bars over three random support-set samplings, and (iii) a table confirming that all baselines were re-evaluated under the identical preprocessing pipeline used for Open-V. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper describes a training-free GFSS method that coordinates frozen external models (SAM3 PCS and K-shot CLIP centroids) via per-pixel arbitration, with no equations, derivations, or self-citations presented that reduce the claimed mIoU gains to fitted parameters, self-defined quantities, or prior author results by construction. The central claim rests on inference-time use of independent foundation-model priors and support examples rather than internal fitting or uniqueness theorems imported from self-citations. No load-bearing steps match the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the assumption that frozen foundation models already encode usable open-vocabulary segmentation priors that can be arbitrated; no free parameters are explicitly named but the calibration step implies at least one tunable scalar or weighting rule; no new entities are introduced.

free parameters (1)
  • arbitration calibration parameter
    Mentioned as part of calibrated per-pixel semantic arbitration; value or selection procedure not specified in abstract.
axioms (1)
  • domain assumption Frozen SAM and CLIP models possess strong open-vocabulary recognition and segmentation capabilities that remain effective on label-disjoint vocabularies.
    Invoked to justify that support information can be incorporated through inference-time grounding without training.

pith-pipeline@v0.9.1-grok · 5807 in / 1422 out tokens · 18962 ms · 2026-06-27T17:10:58.754555+00:00 · methodology

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 · 3 canonical work pages · 1 internal anchor

  1. [1]

    Fossil: Free Open-V ocabulary Semantic Segmentation through Synthetic References Retrieval

    Luca Barsellotti, Roberto Amoroso, Lorenzo Baraldi, and Rita Cucchiara. Fossil: Free Open-V ocabulary Semantic Segmentation through Synthetic References Retrieval. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1464–1473, 2024. 3

    2024

  2. [2]

    Training-Free Open-V ocabulary Segmentation with Offline Diffusion- Augmented Prototype Generation

    Luca Barsellotti, Roberto Amoroso, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Training-Free Open-V ocabulary Segmentation with Offline Diffusion- Augmented Prototype Generation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3689–3698, 2024. 3

    2024

  3. [3]

    SAM 3: Segment Anything with Concepts

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoub- hik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. SAM 3: Segment Anything with Concepts. InIn- ternational Conference on Learning Representations (ICLR),

  4. [4]

    Training-free fine-grained semantic segmentations in low data regimes: A fungitastic baseline

    Sebastian Cavada, Francesco Pelosin, and Lapo Faggi. Training-free fine-grained semantic segmentations in low data regimes: A fungitastic baseline. 2026. Accepted at the 13th Workshop on Fine-Grained Visual Categorization, CVPR 2026. 3

    2026

  5. [5]

    Learning to Generate Text-Grounded Mask for Open-World Semantic Segmentation from Only Image-Text Pairs

    Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to Generate Text-Grounded Mask for Open-World Semantic Segmentation from Only Image-Text Pairs. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11165–11174, 2023. 3

    2023

  6. [6]

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs.IEEE Transac- tions on Pattern Analysis and Machine Intelligence (TPAMI),

  7. [7]

    CAT-Seg: 8 Cost Aggregation for Open-V ocabulary Semantic Segmen- tation

    Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. CAT-Seg: 8 Cost Aggregation for Open-V ocabulary Semantic Segmen- tation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4113–4123, 2024. 3

    2024

  8. [8]

    Open- V ocabulary Universal Image Segmentation with MaskCLIP

    Zheng Ding, Jieke Wang, and Zhuowen Tu. Open- V ocabulary Universal Image Segmentation with MaskCLIP. InInternational Conference on Learning Representations (ICLR), 2023. 3

    2023

  9. [9]

    Self- Support Few-Shot Semantic Segmentation

    Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self- Support Few-Shot Semantic Segmentation. InEuropean Conference on Computer Vision (ECCV), 2022. 3

    2022

  10. [10]

    Enhancing Generalized Few-Shot Se- mantic Segmentation via Effective Knowledge Transfer

    Xinwei Geng et al. Enhancing Generalized Few-Shot Se- mantic Segmentation via Effective Knowledge Transfer. In AAAI Conference on Artificial Intelligence (AAAI), 2024. 1, 3

    2024

  11. [11]

    Visual Prompting for Gen- eralized Few-Shot Segmentation: A Multi-Scale Approach

    Mir Rayat Imtiaz Hossain et al. Visual Prompting for Gen- eralized Few-Shot Segmentation: A Multi-Scale Approach. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3, 5, 6

    2024

  12. [12]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment Anything. InIEEE/CVF Interna- tional Conference on Computer Vision (ICCV), 2023. 1, 3

    2023

  13. [13]

    Learning What Not to Segment: A New Perspective on Few- Shot Segmentation

    Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning What Not to Segment: A New Perspective on Few- Shot Segmentation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  14. [14]

    Open-V ocabulary Semantic Segmentation with Mask-Adapted CLIP

    Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-V ocabulary Semantic Segmentation with Mask-Adapted CLIP. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2023. 3

    2023

  15. [15]

    Class-Agnostic Few-Shot Object Counting and Segmentation via Cross-Attention Networks

    Lan Liu et al. Class-Agnostic Few-Shot Object Counting and Segmentation via Cross-Attention Networks. InAAAI Conference on Artificial Intelligence (AAAI), 2023. 3

    2023

  16. [16]

    Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching

    Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching. InIn- ternational Conference on Learning Representations (ICLR),

  17. [17]

    Intermediate Prototype Mining Trans- former for Few-Shot Semantic Segmentation

    Yuanwei Liu et al. Intermediate Prototype Mining Trans- former for Few-Shot Semantic Segmentation. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. 4

    2022

  18. [18]

    SegCLIP: Patch Aggregation with Learnable Centers for Open-V ocabulary Semantic Segmentation

    Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. SegCLIP: Patch Aggregation with Learnable Centers for Open-V ocabulary Semantic Segmentation. InIn- ternational Conference on Machine Learning (ICML), pages 23033–23044, 2023. 3

    2023

  19. [19]

    Hypercorre- lation Squeeze for Few-Shot Segmentation

    Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorre- lation Squeeze for Few-Shot Segmentation. InIEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3

    2021

  20. [20]

    V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning Robust Visual Features Without Supervi- sion.Transactions on Machine Learning Research (TMLR),

  21. [21]

    Hierarchical Dense Correlation Distillation for Few-Shot Segmentation

    Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical Dense Correlation Distillation for Few-Shot Segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3

    2023

  22. [22]

    Learning Transferable Visual Models from Natural Language Supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models from Natural Language Supervision. InInterna- tional Conference on Machine Learning (ICML), 2021. 1, 3

    2021

  23. [23]

    SAM 2: Seg- ment Anything in Images and Videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Seg- ment Anything in Images and Videos. InInternational Con- ference on Learning Representations (ICLR), 2025. 1, 3

    2025

  24. [24]

    LPOSS: Label Propagation Over Patches and Pixels for Open-V ocabulary Semantic Segmentation.arXiv preprint arXiv:2503.19777, 2025

    Vladan Stojni ´c, Yannis Kalantidis, Jiri Matas, and Giorgos Tolias. LPOSS: Label Propagation Over Patches and Pixels for Open-V ocabulary Semantic Segmentation.arXiv preprint arXiv:2503.19777, 2025. 1, 3

    work page arXiv 2025

  25. [25]

    VRP-SAM: SAM with Visual Reference Prompt

    Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. VRP-SAM: SAM with Visual Reference Prompt. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 8

    2024

  26. [26]

    Prior Guided Feature Enrich- ment Network for Few-Shot Segmentation.IEEE Transac- tions on Pattern Analysis and Machine Intelligence (TPAMI), 44(2):1050–1065, 2020

    Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior Guided Feature Enrich- ment Network for Few-Shot Segmentation.IEEE Transac- tions on Pattern Analysis and Machine Intelligence (TPAMI), 44(2):1050–1065, 2020. 2, 3

    2020

  27. [27]

    Generalized Few-Shot Se- mantic Segmentation

    Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Generalized Few-Shot Se- mantic Segmentation. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2022. 1, 3, 5, 6

    2022

  28. [28]

    Few-Shot Semantic Segmentation Meets SAM3

    Yi-Jen Tsai, Yen-Yu Lin, and Chien-Yao Wang. Few- Shot Semantic Segmentation Meets SAM3.arXiv preprint arXiv:2604.05433, 2026. 3, 7

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  29. [29]

    PANet: Few-Shot Image Semantic Segmen- tation with Prototype Alignment

    Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. PANet: Few-Shot Image Semantic Segmen- tation with Prototype Alignment. InIEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2

    2019

  30. [30]

    Focus on query: Adversarial mining transformer for few-shot segmen- tation.Advances in neural information processing systems, 36:31524–31542, 2023

    Yuan Wang, Naisong Luo, and Tianzhu Zhang. Focus on query: Adversarial mining transformer for few-shot segmen- tation.Advances in neural information processing systems, 36:31524–31542, 2023. 1

    2023

  31. [31]

    Make it up: Fake images, real gains in generalized few-shot semantic segmentation.arXiv preprint arXiv:2603.27206, 2026

    Guohuan Xie, Xin He, Dingying Fan, Le Zhang, Ming-Ming Cheng, and Yun Liu. Make it up: Fake images, real gains in generalized few-shot semantic segmentation.arXiv preprint arXiv:2603.27206, 2026. 3

    work page arXiv 2026

  32. [32]

    A Simple Baseline for Open-V ocabulary Semantic Segmentation with Pre-Trained Vision-Language Model

    Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A Simple Baseline for Open-V ocabulary Semantic Segmentation with Pre-Trained Vision-Language Model. InEuropean Conference on Com- puter Vision (ECCV), pages 736–753, 2022. 3 9

    2022

  33. [33]

    Side Adapter Network for Open-V ocabulary Se- mantic Segmentation

    Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi- ang Bai. Side Adapter Network for Open-V ocabulary Se- mantic Segmentation. InIEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 2945– 2954, 2023. 3

    2023

  34. [34]

    Self-Calibrated Cross Attention Network for Few- Shot Segmentation

    Qianxiong Xu, Wenting Zhao, Guosheng Lin, and Cheng Long. Self-Calibrated Cross Attention Network for Few- Shot Segmentation. InIEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3

    2023

  35. [35]

    Eliminating Feature Ambi- guity for Few-Shot Segmentation

    Qianxiong Xu, Guosheng Lin, Chen Change Loy, Cheng Long, Ziyue Li, and Rui Zhao. Eliminating Feature Ambi- guity for Few-Shot Segmentation. InEuropean Conference on Computer Vision (ECCV), 2024. 2

    2024

  36. [36]

    Unlocking the Power of SAM 2 for Few-Shot Segmentation

    Qianxiong Xu, Lanyun Zhu, Xuanyi Liu, Guosheng Lin, Cheng Long, Ziyue Li, and Rui Zhao. Unlocking the Power of SAM 2 for Few-Shot Segmentation. InInternational Con- ference on Machine Learning (ICML), 2025. 3

    2025

  37. [37]

    Feature- Proxy Transformer for Few-Shot Segmentation

    Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Feature- Proxy Transformer for Few-Shot Segmentation. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. 3

    2022

  38. [38]

    Open V ocabulary Scene Parsing

    Hang Zhao, Xavier Puig, Bolei Zhou, Sanja Fidler, and Antonio Torralba. Open V ocabulary Scene Parsing. In IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 3

    2017

  39. [39]

    predictor–loader transform equiv- alence

    Yaoxin Zhuo, Zachary Bessinger, Lichen Wang, Naji Khos- ravan, Baoxin Li, and Sing Bing Kang. TFM2: Training-Free Mask Matching for Open-V ocabulary Semantic Segmenta- tion. InIEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 1, 3 10 A. Alignment-Diagnostic Reference Imple- mentation This appendix supports the methodological reco...

    2025

  40. [40]

    For each candidate ADE150 class, split its name on “;” into synonyms (e.g.streetlight;street lamp→ {streetlight, street lamp})

  41. [41]

    For each synonym, collect every WordNet noun-synset lemma reachable from any of its tokens (no hyper- nym/hyponym expansion)

  42. [42]

    Form the union of these lemma sets over all synonyms of the candidate; call itL c

  43. [43]

    VOC and IN-1k labels are kept whole, so table lampdoesnotleak the bare classlamp

    Build the same lemma sets for every COCO- stuff, VOC, and IN-1k label, with one prepro- cessing exception: COCO-stuff compound labels of the form{class}-other,{class}-stuff, {class}-mergedare decomposed into their whites- pace tokens (e.g.sky-othercontributes the bare lemmasky). VOC and IN-1k labels are kept whole, so table lampdoesnotleak the bare classlamp

  44. [44]

    val n” / “train n

    Drop the candidate iffL c intersects any of these three union lemma sets. The full filter is implemented in owgfss/benchmarks/ade ow/build ade ow.py and runs in seconds on CPU; the script writes the kept and rejected lists toade ow classes.jsonand ade ow rejected.jsonrespectively. B.2. Retained classes Table S2 lists the 26 surviving ADE150 classes with t...

    1989