Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

OV-COAST: Cost Aggregation with Optimal Transport for Open-Vocabulary Semantic Segmentation

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Adding optimal transport alignment to CAT-Seg's cost volume lifts open-vocabulary segmentation by 1.72% mIoU on the MESS benchmark.

desk verdict A reasonable empirical idea—adding Sinkhorn OT to CAT-Seg's cost volume—but the paper never specifies how the frozen transport plan affects training, so the headline gain is not verifiable as written. read the letter →

arxiv 2506.03706 v1 pith:COW33V66 submitted 2025-06-04 cs.CV

Aditya Gandhamal , Aniruddh Sikdar , Suresh Sundaram This is my paper
classification cs.CV
keywords open-vocabularysemanticsegmentationoptimaltransportcostaggregationSinkhorndistanceCLIPzero-shotgeneralizationMESSbenchmarktwo-stageoptimization
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

This paper proposes OV-COAST, a modification of the CAT-Seg open-vocabulary segmentation model that inserts an optimal transport alignment step. The idea is to treat the multi-modal cost volume, which measures cosine similarity between CLIP image and text embeddings, as a cost matrix and solve an optimal transport problem with the Sinkhorn algorithm. The resulting transport plan, computed with frozen features and through repeated sampling, is then frozen while CAT-Seg is trained on COCO-Stuff with cross-entropy loss. On the MESS benchmark of 19 out-of-domain datasets, the method reports a mean IoU of 31.15, surpassing CAT-Seg by 1.72% and SAN-B by 4.94%. The paper argues that this two-stage alignment improves multi-modal feature alignment and out-of-domain generalization.

What carries the argument

The load-bearing object is the cost volume from CAT-Seg, repurposed as the cost matrix of an optimal transport problem. Following PLOT, the paper defines discrete distributions over visual features $D_V$ and textual features $D_L$, sets the cost as $C_{m,n} = 1 - \text{Caggr}\{i,n\}$, and computes the transport plan via the Sinkhorn iteration $T^* = \operatorname{diag}(u^{(t)}) \exp(-C/\lambda) \operatorname{diag}(v^{(t)})$. The plan is computed with frozen encoders in an inner loop and then kept fixed in the outer loop while the CAT-Seg model is trained, which is the mechanism intended to align visual and textual features before the segmentation head is learned.

What would settle it

Train CAT-Seg with the same settings but replace the frozen OT transport plan with a random matrix of the same shape, or simply remove the OT stage entirely; if the MESS mIoU does not drop substantially relative to 31.15, the reported gain is not attributable to the optimal transport alignment.

Watch

Extended reading notes

Core claim

The central claim is that solving an optimal transport problem on CAT-Seg's cost volume improves open-vocabulary semantic segmentation under domain shift. The paper constructs the OT cost matrix as $C_{m,n} = 1 - \text{Caggr}\{i,n\}$, where $\text{Caggr}$ is the cosine-similarity cost volume between CLIP image embeddings $D_V$ and text embeddings $D_L$. In the first stage, the transport plan $T^*$ is computed by minimizing $\langle T, C \rangle$ with the Sinkhorn distance while keeping the CLIP encoders fixed; in the second stage, the plan is frozen and the segmentation model is trained with cross-entropy following CAT-Seg's strategy. The reported result is that this addition raises MESS benchmark mIoU from 29.42 (CAT-Seg) to 31.15 (OV-COAST), with notable gains on CHASE DB1, Corrosion CS, and DeepCrack. The paper also reports that adding prompt-learning methods on top of this OT step hurts performance, falling below CAT-Seg.

Load-bearing premise

The central assumption is that a transport plan computed from the cost volume with frozen features, and then frozen during training, actually guides the segmentation model to better generalization; the paper does not specify any mechanism by which the fixed plan influences the cross-entropy training.

Editorial extensions

If this is right

  • If the reported gains hold, OV-COAST establishes that optimal transport alignment on the cost volume is a lightweight plug-in that improves CAT-Seg's out-of-domain robustness without changing its architecture or inference cost.
  • The method suggests that freezing a pre-computed alignment plan is sufficient to guide segmentation training, so the expensive OT solve can be done once and reused across training runs.
  • The ablation result implies that combining OT alignment with prompt-learning techniques such as PLOT or CoPLOT is counterproductive for this pipeline, pointing to a trade-off between prompt adaptation and transport-based alignment.
  • With a ViT-B backbone only, the gains leave open whether the same margin persists with larger backbones or other cost-aggregation baselines.

Reading between the lines

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

  • The paper does not specify how the frozen transport plan enters the loss or the forward pass during segmentation training; if the plan is used only as an initialization or a regularizer, it may be replaceable by a cheaper heuristic such as a fixed random or identity plan, and testing that replacement would isolate whether OT is the cause of the gain.
  • The reported per-dataset gains are uneven, with large jumps on CHASE DB1 (16.99 to 30.30) and Corrosion CS (12.29 to 14.47) but decreases on FoodSeg103 and ATLANTIS; this suggests the OT alignment may help most where text-to-visual correspondence is weak, which could be tested on datasets with known low CLIP alignment.
  • Because the method inherits CAT-Seg's two-stage spatial and class aggregation, a natural extension is to let the transport plan be updated during training rather than frozen, which would test whether the frozen plan is a feature or a limitation.
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 / 4 minor

Summary. The paper proposes OV-COAST, an extension of CAT-Seg for open-vocabulary semantic segmentation. It constructs a multi-modal cost volume from CLIP image and text embeddings, defines an optimal transport cost matrix as one minus the cosine-similarity cost volume, solves for a transport plan T* via Sinkhorn iterations, and claims a two-stage training procedure in which the frozen transport plan guides CAT-Seg training. The paper reports a mean mIoU of 31.15 on the MESS benchmark versus 29.42 for CAT-Seg and 26.21 for SAN-B, and it compares with several prompt-learning-plus-OT variants in Table 3.

Significance. Conditional on the claimed mechanism, augmenting CAT-Seg with an OT-based alignment would be a simple and potentially generalizable addition to a strong OVSS baseline, and the MESS benchmark is a relevant test bed for out-of-domain generalization. The paper's code release is a positive but not sufficient feature. However, the contribution is not verifiable from the text: the central training mechanism is undefined, the reported gains lack any variance or significance analysis, and the closely related ZegOT baseline is omitted. The significance of the empirical results therefore cannot be assessed as the manuscript stands.

major comments (4)
  1. [Section 4.1, around Eq. (5)] The paper never specifies how the frozen transport plan T* enters the outer-loop training objective. Section 4.1 says T* "remains fixed while the model is trained for the OVSS task using cross-entropy loss," but the cross-entropy loss is a function of the model predictions; without an explicit loss term such as L = L_CE + λ<T*, C>, or an architectural injection of T* into the cost-aggregation modules, the gradient of the loss with respect to the model parameters is independent of T*. Consequently the claimed 1.72% improvement over CAT-Seg in Table 2 is not attributable to the proposed OT component as written. The authors must state the exact loss, the forward path in which T* is used, and how gradients flow (or do not flow) through it.
  2. [Table 2 and Section 5] The reported mIoU values are single runs with no error bars, multiple seeds, or significance testing. The per-dataset pattern is not uniformly positive: OV-COAST is below CAT-Seg on FoodSeg103 (25.92 vs 26.69), ATLANTIS (39.59 vs 40.31), and Kvasir-inst. (46.56 vs 48.2), among others. Given that the mean gap is 1.72 points, the result could plausibly fall within run-to-run variability. Please provide standard deviations over at least three seeds and a paired comparison across the 19 datasets.
  3. [Related Works and Section 5, Table 3] The paper cites ZegOT [12], which also uses optimal transport for zero-shot segmentation, but it never compares against ZegOT as a baseline. Table 3 evaluates CAT-Seg with VPT, CoPLOT, and PLOT, but not with ZegOT. Since ZegOT is the closest OT-based baseline, its inclusion is necessary to isolate whether any improvement comes from OT alignment rather than from incidental differences in training setup or evaluation.
  4. [Section 4.1, Eq. (6)] Because T* is computed as a deterministic function of the same cost volume (C = 1 - Caggr) that CAT-Seg already uses for prediction, it is unclear what independent information the frozen plan provides. The paper offers no analysis or ablation showing that T* carries signal beyond the cost volume; for example, replacing T* with a uniform or identity plan in the same training recipe would test whether the gains are due to OT specifically. As written, the risk that the alignment step merely reinforces existing similarities in the cost volume is not addressed.
minor comments (4)
  1. [Abstract] The code URL contains a duplicated brace artifact: 'https://github.com/adityagandhamal/OV-COAST/}{https://github.com/adityagandhamal/OV-COAST/ .' Please fix the formatting.
  2. [Section 3, Eq. (1)] The notation δDV and δDL in Eq. (1) is not defined; the subscript appears to denote a feature value rather than a distribution. Please clarify the measure-theoretic notation.
  3. [Table 3] Table 3 reports mIoU on MESS but does not include the plain CAT-Seg baseline in the same table, so the claim that prompt-learning variants "even fall below CAT-Seg" cannot be verified without switching to Table 2. Please add the baseline row or explicitly cross-reference Table 2.
  4. [Section 5] The text says the MESS benchmark contains 22 datasets but the model is tested on 19; please explain which 3 datasets are excluded and why.

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrated circularity; the paper's two-stage OT guidance is under-specified rather than self-referential, so the claimed gain over CAT-Seg is unverifiable but not circular by construction.

full rationale

The paper's central claim is that optimal-transport (OT) guidance from the cost volume improves CAT-Seg on the MESS benchmark. A circularity finding requires exhibiting a specific reduction: e.g., an equation that defines the prediction in terms of the fitted input, or a fitted parameter renamed as a prediction. Here, the method defines the OT cost matrix as Cm,n = (1 - Caggr{i,n}) from the same cost volume used by CAT-Seg (Eq. 6), and then solves for T* via Sinkhorn (Eq. 5). However, the paper never states how T* enters the second-stage cross-entropy loss, the forward pass, or the gradients. Section 4.1 says only that 'the transport plan T* remains fixed while the model is trained for the OVSS task using cross-entropy loss,' with no formula or architectural injection specified. Consequently, one cannot exhibit a derivation in which the reported mIoU reduces to the fitted T* by construction. The most serious issue is incompleteness: the improvement over CAT-Seg is unsupported as written because the mechanism connecting T* to the training objective is missing, and no error bars or multiple-seed results are reported. That is a verifiability and correctness problem, not a circularity problem under the stated hard rules. The references to the authors' prior work (Refs. 21-25) are not load-bearing for the OT method, and no uniqueness theorem or ansatz is smuggled in via self-citation. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim relies on the assumption that OT-based alignment of the same features used by CAT-Seg improves generalization. The method section omits how the transport plan is used in the loss, so the ledger is small but the missing mechanism is the main risk.

free parameters (2)
  • Sinkhorn regularization lambda
    The regularization parameter in Eq. (5) is not specified in the paper; it controls the entropic regularization of the transport plan.
  • Convergence threshold Delta_v = 0.01
    Set to 0.01 in Section 5, the inner loop stopping criterion for the transport plan update; no sensitivity analysis is given.
assumptions (3)
  • standard math Sinkhorn's algorithm converges to the entropic OT solution for the given cost matrix
    Invoked in Sections 3 and 4.1 without proof, standard result.
  • domain assumption Aligning visual and textual feature distributions via OT improves zero-shot segmentation generalization
    The paper's core premise, stated in Sections 1 and 4, is that OT-based alignment enhances out-of-domain performance; no independent evidence is provided for this assumption beyond the reported results.
  • ad hoc to paper The transport plan T* can be frozen and used to guide CAT-Seg training without altering the model architecture
    Asserted in Section 4.1, but no equation links T* to the cross-entropy loss or prediction head.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OV-COAST: Cost Aggregation with Optimal Transport for Open-Vocabulary Semantic Segmentation." pith.science (2026). https://pith.science/paper/COW33V66

@misc{pith2026250603706,
  author       = {Pith},
  title        = {Pith review of: OV-COAST: Cost Aggregation with Optimal Transport for Open-Vocabulary Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/COW33V66}},
  note         = {Machine review of arXiv:2506.03706}
}
read the original abstract

Open-vocabulary semantic segmentation (OVSS) entails assigning semantic labels to each pixel in an image using textual descriptions, typically leveraging world models such as CLIP. To enhance out-of-domain generalization, we propose Cost Aggregation with Optimal Transport (OV-COAST) for open-vocabulary semantic segmentation. To align visual-language features within the framework of optimal transport theory, we employ cost volume to construct a cost matrix, which quantifies the distance between two distributions. Our approach adopts a two-stage optimization strategy: in the first stage, the optimal transport problem is solved using cost volume via Sinkhorn distance to obtain an alignment solution; in the second stage, this solution is used to guide the training of the CAT-Seg model. We evaluate state-of-the-art OVSS models on the MESS benchmark, where our approach notably improves the performance of the cost-aggregation model CAT-Seg with ViT-B backbone, achieving superior results, surpassing CAT-Seg by 1.72 % and SAN-B by 4.9 % mIoU. The code is available at https://github.com/adityagandhamal/OV-COAST/}{https://github.com/adityagandhamal/OV-COAST/ .

Figures

Figures reproduced from arXiv: 2506.03706 by the authors.

Figure 1
Figure 1. Visualization of logits of the vision and text embed [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of OV-COAST. We compute cost from image [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of CAT-Seg and OV-COAST(ours) on MHP, CWFID and CHASE datasets from the MESS benchmark. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. DINOde: Continuous Vision-Text Alignment for Open-Vocabulary Semantic Segmentation

    cs.CV 2026-07 conditional novelty 5.0 of 10

    A 10-step ODE text-to-vision flow with tangent-space projection outperforms single-step MLP projection for open-vocabulary semantic segmentation, reaching 49.5 average mIoU without mask refinement.

Reference graph

Works this paper leans on

36 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [12]

    ZegOT: Zero-shot Segmentation Through Optimal Transport of Text Prompts

    Kwanyoung Kim, Yujin Oh, and Jong Chul Ye. Zegot: Zero- shot segmentation through optimal transport of text prompts. arXiv preprint arXiv:2301.12171, 2023. 2

    work page Pith review arXiv 2023

  2. [1]

    What a mess: Multi-domain evaluation of zero-shot semantic segmentation

    Benedikt Blumenstiel, Johannes Jakubik, Hilde K ¨uhne, and Michael V¨ossing. What a mess: Multi-domain evaluation of zero-shot semantic segmentation. Advances in Neural Infor- mation Processing Systems, 36, 2024. 1, 2

    work page 2024

  3. [2]

    Coco- stuff: Thing and stuff classes in context

    Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Coco- stuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recog- nition, pages 1209–1218, 2018. 4

    work page 2018

  4. [3]

    Plot: Prompt learning with optimal transport for vision-language models

    Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. 2, 3

    arXiv 2022

  5. [4]

    Cat- seg: Cost aggregation for open-vocabulary semantic seg- mentation

    Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat- seg: Cost aggregation for open-vocabulary semantic seg- mentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4113– 4123, 2024. 1, 2, 3, 4

    work page 2024

  6. [5]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information pro- cessing systems, 26, 2013. 3

    work page 2013

  7. [6]

    De- coupling zero-shot semantic segmentation

    Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. De- coupling zero-shot semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583–11592, 2022. 1, 2

    work page 2022

  8. [7]

    Open- vocabulary panoptic segmentation maskclip

    Zheng Ding, Jieke Wang, and Zhuowen Tu. Open- vocabulary panoptic segmentation maskclip. 2022. 2

    work page 2022

Show all 36 references
  1. [8]

    The pascal visual object classes challenge: A retrospective

    Mark Everingham, SM Ali Eslami, Luc Van Gool, Christo- pher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. In- ternational journal of computer vision , 111:98–136, 2015. 2

    2015

  2. [9]

    Scal- ing open-vocabulary image segmentation with image-level labels

    Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scal- ing open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540–557. Springer, 2022. 1, 2

    2022

  3. [10]

    Scaling up visual and vision-language representa- tion learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representa- tion learning with noisy text supervision. In International conference on machine learning, pages 4904–4916. PMLR,

  4. [11]

    End-to-end learning of geometry and context for deep stereo regression

    Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Proceedings of the IEEE international confer- ence on computer vision, pages 66–75, 2017. 4

    2017

  5. [13]

    Fine-tuning can distort pretrained fea- tures and underperform out-of-distribution

    Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained fea- tures and underperform out-of-distribution. arXiv preprint arXiv:2202.10054, 2022. 1

    2022 arXiv

  6. [14]

    Open-vocabulary 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-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pag...

    2023

  7. [15]

    Graftnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature

    Biyang Liu, Huimin Yu, and Guodong Qi. Graftnet: Towards domain generalized stereo matching with a broad-spectrum and task-oriented feature. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 13012–13021, 2022. 2

    2022

  8. [16]

    M ´emoire sur la th ´eorie des d ´eblais et des remblais

    Gaspard Monge. M ´emoire sur la th ´eorie des d ´eblais et des remblais. Mem. Math. Phys. Acad. Royale Sci. , pages 666– 704, 1781. 2, 3

  9. [17]

    Computational optimal transport: With applications to data science

    Gabriel Peyr ´e, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355–607, 2019. 2

    2019

  10. [18]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  11. [19]

    Convo- lutional neural network architecture for geometric matching

    Ignacio Rocco, Relja Arandjelovic, and Josef Sivic. Convo- lutional neural network architecture for geometric matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6148–6157, 2017. 2, 4

    2017

  12. [20]

    Lca-on-the-line: Benchmarking out-of-distribution generalization with class taxonomies

    Jia Shi, Gautam Gare, Jinjin Tian, Siqi Chai, Zhiqiu Lin, Arun Vasudevan, Di Feng, Francesco Ferroni, and Shu Kong. Lca-on-the-line: Benchmarking out-of-distribution generalization with class taxonomies. arXiv preprint arXiv:2407.16067, 2024. 1

    2024 arXiv

  13. [21]

    Fully complex-valued fully con- volutional multi-feature fusion network (fc 2 mfn) for build- ing segmentation of insar images

    Aniruddh Sikdar, Sumanth Udupa, Suresh Sundaram, and Narasimhan Sundararajan. Fully complex-valued fully con- volutional multi-feature fusion network (fc 2 mfn) for build- ing segmentation of insar images. In 2022 IEEE Symposium Series on Computational Intelligence (SSCI) , pages 581–

    2022

  14. [22]

    Deepmao: Deep multi-scale aware over- complete network for building segmentation in satellite im- agery

    Aniruddh Sikdar, Sumanth Udupa, Prajwal Gurunath, and Suresh Sundaram. Deepmao: Deep multi-scale aware over- complete network for building segmentation in satellite im- agery. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 487–496, 2023

    2023

  15. [23]

    Fully complex-valued deep learning model for visual percep- tion

    Aniruddh Sikdar, Sumanth Udupa, and Suresh Sundaram. Fully complex-valued deep learning model for visual percep- tion. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

    2023

  16. [24]

    Ogp- net: Optical guidance meets pixel-level contrastive distilla- tion for robust multi-modal and missing modality segmen- tation

    Aniruddh Sikdar, Jayant Teotia, and Suresh Sundaram. Ogp- net: Optical guidance meets pixel-level contrastive distilla- tion for robust multi-modal and missing modality segmen- tation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6922–6930, 2025

    2025

  17. [25]

    Mrfp: Learning generalizable semantic segmentation from sim-2-real with multi-resolution feature perturbation

    Sumanth Udupa, Prajwal Gurunath, Aniruddh Sikdar, and Suresh Sundaram. Mrfp: Learning generalizable semantic segmentation from sim-2-real with multi-resolution feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 59...

    2024

  18. [26]

    Robust fine-tuning of zero-shot models

    Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gon- tijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vi- ...

    2022

  19. [27]

    Open-vocabulary panop- tic segmentation with text-to-image diffusion models

    Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiao- long Wang, and Shalini De Mello. Open-vocabulary panop- tic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 2955–2966, 2023. 2

    2023

  20. [28]

    A simple baseline for open- vocabulary 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- vocabulary semantic segmentation with pre-trained vision- language model. In European Conference on Computer Vi- sion, pages 736–753. Springer, 2022. 1, 2

    2022

  21. [29]

    Side adapter network for open-vocabulary semantic segmentation

    Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi- ang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2945– 2954, 2023. 2

    2023

  22. [30]

    Convolutions die hard: Open-vocabulary seg- mentation with single frozen convolutional clip

    Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang- Chieh Chen. Convolutions die hard: Open-vocabulary seg- mentation with single frozen convolutional clip. Advances in Neural Information Processing Systems, 36:32215–32234,

  23. [31]

    Florence: A new foundation model for computer vision

    Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1

    2021 arXiv

  24. [32]

    Dept: Decoupled prompt tuning

    Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924–12933, 2024. 2

    2024

  25. [33]

    Semantic under- standing of scenes through the ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fi- dler, Adela Barriuso, and Antonio Torralba. Semantic under- standing of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302–321, 2019. 2

    2019

  26. [34]

    Extract free dense labels from clip

    Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Com- puter Vision, pages 696–712. Springer, 2022. 1

    2022

  27. [35]

    Learning to prompt for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In- ternational Journal of Computer Vision, 130(9):2337–2348,

  28. [36]

    Zegclip: Towards adapting clip for zero-shot seman- tic segmentation

    Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yi- fan Liu. Zegclip: Towards adapting clip for zero-shot seman- tic segmentation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.