Pith. sign in

REVIEW 4 major objections 5 minor 4 cited by

Seg-R1: Segmentation Can Be Surprisingly Simple with Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Pure reinforcement learning, with no text supervision and no architectural changes, can turn a general vision-language model into a segmenter that transfers zero-shot to referring and reasoning segmentation.

desk verdict Clever RL-prompting recipe for segmentation, but the zero-shot transfer claim needs a base-model control before it convinces. read the letter →

arxiv 2506.22624 v1 pith:MJFJIVCF submitted 2025-06-27 cs.CV

classification cs.CV
keywords reinforcementlearningsegmentationvision-languagemodelGRPOzero-shotgeneralizationcamouflagedobjectdetectionreferringreasoning
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

Seg-R1 claims that pixel-level segmentation can be learned by a large multimodal model through pure reinforcement learning, without text supervision, special segmentation tokens, or architectural changes. The model is trained with GRPO to output a reasoning trace and then point and bounding-box prompts, which a frozen SAM2 converts into masks; the reward combines format compliance with a weighted IoU and S-measure score. On only 7,040 foreground image-mask pairs, the paper reports an S-measure of .873 on COD10K, and zero-shot cIoU of 71.4 on RefCOCOg and gIoU of 56.7 on ReasonSeg, above fully supervised models on those benchmarks. The significance would be a much cheaper route to pixel-level grounding, and evidence that RL can generalize beyond its training task.

What carries the argument

The load-bearing mechanism is GRPO with a two-part reward, applied to a prompt-generation policy. GRPO samples a group of outputs for each image-query pair and computes each output's advantage from its reward relative to the group average, then updates the policy with a clipped importance ratio and a KL penalty to the reference policy. The reward is a hard format check that the output uses the required <think>, <bbox>, <points>, and <labels> tags plus a segmentation reward of 0.7*IoU + 0.3*S-measure; the paper reports that S-measure alone collapses to black masks, and that the combination prevents that reward hacking. The other load-bearing piece is the decomposition: the LMM only predicts sparse prompts, and frozen SAM2 does dense mask generation, so the hard pixel-level supervision signal is replaced by an RL signal on mask quality.

What would settle it

Run the same prompt templates and the same frozen SAM2 decoding on Qwen-2.5-VL with no RL or SFT training, and evaluate on RefCOCOg and ReasonSeg; if the base model already scores near Seg-R1's 71.4 cIoU and 56.7 gIoU, the RL attribution is unsupported. A complementary check is to train with a reward that ignores mask quality, using only the format reward, and see whether the zero-shot transfer disappears.

Watch

Extended reading notes

Core claim

The central claim is that GRPO, applied to mask-prompt prediction, is enough to give an LMM segmentation ability, and that this ability transfers to unseen tasks. Starting from Qwen-2.5-VL and frozen SAM2, Seg-R1 treats segmentation as next-token prediction of prompts; the policy generates the reasoning and the prompts, and the reward module grades the resulting mask. With pure RL on DIS5K, COD10K, and CAMO, using 7,040 image-mask pairs and no textual supervision, the model reaches state-of-the-art COD and SOD numbers, and zero-shot referring and reasoning segmentation results comparable to or better than fully supervised systems. The authors also argue that RL preserves general multimodal performance, whereas the SFT version using their FCoT dataset loses it.

Load-bearing premise

The load-bearing premise is that the zero-shot referring and reasoning segmentation measured after RL reflects what RL taught the model, not what the base Qwen-2.5-VL could already do when prompted the same way; the paper provides no baseline score for the untrained base model on those benchmarks.

Editorial extensions

If this is right

  • If correct, segmentation for LMMs no longer requires large pixel-level image-text corpora or specialized decoder heads; a compact RL stage on about seven thousand masks can produce competitive foreground segmentation.
  • The zero-shot results imply that a model trained only to segment salient or camouflaged foreground objects can acquire generalizable referring and reasoning segmentation skills from the RL signal alone.
  • Because RL preserves scores on general VLM benchmarks while SFT degrades them, the paper points to RL as a safer adaptation method for adding new capabilities to multimodal models.
  • Fine-tuning the same recipe on DUTS lifts Seg-R1 to state-of-the-art salient object detection, suggesting the method transfers across foreground segmentation tasks.

Reading between the lines

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

  • Editorial inference: the paper never reports a no-RL control on RefCOCOg and ReasonSeg, so part of the zero-shot transfer may be inherited from Qwen-2.5-VL's pretrained grounding rather than created by RL; a control experiment would settle this.
  • Editorial inference: the same reward-on-prompt design should apply to other frozen promptable decoders, such as depth or edge estimators, and to other dense prediction tasks, with the caveat that structural metrics may need an IoU-style anchor to avoid reward hacking.
  • Editorial inference: since only 7,040 training pairs were used, scaling behavior is unknown; the approach could saturate or improve non-linearly with more diverse foreground data, and that is a testable extension.
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 presents Seg-R1, a framework that reformulates segmentation as sparse prompt prediction: Qwen-2.5-VL generates bounding box, point, and label tokens, and a frozen SAM2 turns these prompts into masks. The authors train the LMM with GRPO, either after SFT on a new FCoT dataset or in a 'pure RL' pipeline that begins with pre-RL on DIS5K and then RL on COD10K/CAMO. They report an S-measure of .873 on COD10K and strong zero-shot referring and reasoning segmentation (71.4 cIoU on RefCOCOg test and 56.7 gIoU on ReasonSeg test), claiming that pure RL on 7,040 foreground segmentation image-mask pairs without text supervision transfers to these tasks.

Significance. If the attribution is established, the result would be significant: it suggests that sparse prompting with a frozen SAM2 plus GRPO is a scalable, architecture-preserving alternative to special segmentation tokens, and that RL can improve out-of-domain generalization while preserving general VLM capabilities better than SFT. The paper also introduces FCoT as a prompt-annotation dataset, and Figure 7 reports retention on general multimodal benchmarks. A clear strength is the simple next-token prompt formulation and the internal comparison showing that the RL-trained variants outperform the SFT cold-start variant on referring and reasoning benchmarks. However, the headline zero-shot claim currently rests on comparisons against fully supervised models rather than against the base Qwen-2.5-VL pipeline, so the central attribution is untested.

major comments (4)
  1. [§4.3, Tables 3 and 4; Abstract] The zero-shot claim that pure RL produces referring and reasoning segmentation abilities is not supported by the reported comparisons. The tables compare Seg-R1 with fully supervised methods, but no row reports the performance of the base Qwen-2.5-VL (or Qwen-2.5-VL plus the same SAM2 prompt wrapper) on RefCOCO, RefCOCO+, RefCOCOg, and ReasonSeg before RL. Qwen-2.5-VL is a modern instruction-following VLM with strong grounding capabilities, so a large share of the cIoU could pre-exist RL. Table 5's 'baseline' is measured only on CAMO and cannot rule this out. Please add this control and, if the base model already performs well on these benchmarks, revise the attribution from 'equips' to 'fine-tunes'.
  2. [Abstract; §1] The phrases 'purely RL-based training' and 'without any textual supervision' overstate the setup, because the model is initialized from Qwen-2.5-VL, which was trained on large-scale image-text data and likely on referring and grounding data. The RL stage itself may use only image-mask pairs, but the headline as written ignores the text-supervised pretraining of the base model. Please rephrase to say that the RL training stage uses no textual supervision, or provide evidence about the pretraining data.
  3. [§3.1; §4.4, Table 6] The choice of reward weights (0.7 IoU, 0.3 S-measure) appears to be made after observing performance on CAMO, and Table 6 shows that S-measure alone causes reward hacking. The selection procedure is not described. If these weights were tuned on the same benchmark used for the headline COD results, the reported numbers are optimistic. Please state how the weights were selected and whether CAMO served as a validation or test set during development.
  4. [§4.1; Abstract; Conclusion] The claim of '7,040 foreground segmentation image-mask pairs' is ambiguous. Section 4.1 states that the pure-RL pipeline first performs pre-RL on DIS5K-TR with 3,000 images and then RL on COD10K and CAMO. If the 7,040 figure excludes the DIS5K stage, the abstract understates the total data exposure; if it includes DIS5K, the paper should say so explicitly. Please clarify the exact composition of the training data.
minor comments (5)
  1. [§3.2, FCoT] The FCoT dataset relies on Gemini-2.5-Pro to generate chains of thought, but the paper does not validate that these chains correspond to actual human annotator reasoning. This matters for the SFT baseline comparison and should be acknowledged.
  2. [Table 4] In Table 4, Seg-R1-7B reports val cIoU 41.2 but test cIoU 53.7; this is an unusual direction of improvement and should be explained or checked for a typo.
  3. [Table 2] The zero-shot SOD results for Seg-R1-3B on HKU-IS and ECSSD are lower than the SFT cold-start variant; the discussion of zero-shot SOD performance should acknowledge these cases rather than focusing only on the strong DUT-OMRON and DUTS-TE results.
  4. [Figure 7] The general-benchmark comparison in Figure 7 is presented without numerical values; adding a table or numbers would make the claim about preserved general capabilities reproducible.
  5. [§4.1; Project page] The paper does not state whether code or checkpoints will be released; the project page should be supplemented with artifacts if the results are to be independently verified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Seg-R1 is an empirical systems paper with external held-out evaluations; the missing no-RL baseline is a validity concern, not circularity.

full rationale

Seg-R1 is an empirical systems paper: it trains a large multimodal model with GRPO on foreground segmentation datasets and evaluates the resulting model on held-out benchmarks. There is no load-bearing derivation chain whose conclusions are equivalent to its inputs by construction. The reward function combines IoU (0.7) and S-measure (0.3), and the same metrics are later reported on COD10K and CAMO, but this is a standard training-objective/evaluation-metric overlap on separate test splits, not a fitted parameter being renamed as a prediction. The zero-shot referring and reasoning segmentation results are obtained on RefCOCOg and ReasonSeg without RL training on those tasks, and the paper does not report what the base Qwen-2.5-VL model achieves on those benchmarks without RL; this is a missing control that weakens the attribution of the zero-shot ability to the RL stage, but it is an experimental validity issue, not circularity. The '7,040 pairs' framing omitting the DIS5K pre-RL stage is likewise a reporting/presentation concern. Self-citations such as FOCUS appear only as comparison baselines and are not load-bearing. No step in the paper reduces to its own inputs by definition, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The model starts from Qwen-2.5-VL, so the base grounding and frozen SAM2 quality are load-bearing assumptions. The reward weights are hand-selected on validation. FCoT is a new dataset resource, not a theoretical entity, and no new physical or conceptual entities are introduced.

free parameters (2)
  • IoU reward weight = 0.7
    Weight of IoU in the combined segmentation reward; selected using CAMO validation (Table 6) with no sensitivity analysis reported.
  • S-measure reward weight = 0.3
    Weight of S-measure in the combined segmentation reward; selected together with the IoU weight on CAMO validation.
assumptions (5)
  • domain assumption Qwen-2.5-VL provides robust vision-language grounding sufficient to interpret natural language expressions.
    Zero-shot referring and reasoning segmentation rely on the base model's language understanding (Sec 4.3).
  • domain assumption SAM2 generates high-quality masks from point and box prompts, including in zero-shot settings.
    Final masks are produced by frozen SAM2, so any mask error upper-bounds Seg-R1; the paper acknowledges SAM2 struggles on camouflaged objects (Sec 4.3).
  • domain assumption GRPO policy optimization on token-level prompts improves prompt quality and transfers to new tasks.
    Taken from prior RL literature and assumed to hold for this segmentation setting (Sec 3.1).
  • domain assumption Ground-truth masks from DIS5K, COD10K, and CAMO are sufficiently accurate to serve as RL rewards.
    The reward is computed against these masks (Sec 4.1).
  • ad hoc to paper FCoT's Gemini-generated chains of thought faithfully describe human annotation reasoning.
    FCoT is used for the SFT baseline, but no human evaluation of the CoT quality is provided (Sec 3.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Seg-R1: Segmentation Can Be Surprisingly Simple with Reinforcement Learning." pith.science (2026). https://pith.science/paper/MJFJIVCF

@misc{pith2026250622624,
  author       = {Pith},
  title        = {Pith review of: Seg-R1: Segmentation Can Be Surprisingly Simple with Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJFJIVCF}},
  note         = {Machine review of arXiv:2506.22624}
}
read the original abstract

We present Seg-R1, a preliminary exploration of using reinforcement learning (RL) to enhance the pixel-level understanding and reasoning capabilities of large multimodal models (LMMs). Starting with foreground segmentation tasks, specifically camouflaged object detection (COD) and salient object detection (SOD), our approach enables the LMM to generate point and bounding box prompts in the next-token fashion, which are then used to guide SAM2 in producing segmentation masks. We introduce Group Relative Policy Optimization (GRPO) into the segmentation domain, equipping the LMM with pixel-level comprehension through a carefully designed training strategy. Notably, Seg-R1 achieves remarkable performance with purely RL-based training, achieving .873 S-measure on COD10K without complex model modification. Moreover, we found that pure RL training demonstrates strong open-world generalization. Despite being trained solely on foreground segmentation image-mask pairs without text supervision, Seg-R1 achieves impressive zero-shot performance on referring segmentation and reasoning segmentation tasks, with 71.4 cIoU on RefCOCOg test and 56.7 gIoU on ReasonSeg test, outperforming models fully supervised on these datasets.

Figures

Figures reproduced from arXiv: 2506.22624 by the authors.

Figure 1
Figure 1. Overview of segmentation ability of Seg-R1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Seg-R1 framework. We introduce GRPO into the segmentation domain, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. A visualization of FCoT, a dataset that captures the step-by-step reasoning process used [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of single object referring segmentation in the wild. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison of multiple objects referring segmentation in the wild. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Actor as Its Own Critic: Unifying Region Understanding and Localization via CycleGRPO

    cs.CV 2026-07 accept novelty 7.0 of 10

    A single MLLM jointly improves region captioning and localization by rewarding captions that let it reconstruct the original mask, needing only region inputs.

  2. Credit the Right Box: Marginal Contribution Assignment for Structured Visual Perception

    cs.CV 2026-08 conditional novelty 6.0 of 10

    Box-level leave-one-out credit, derived from a continuous matched-set score, improves GRPO training of multimodal models for grounding, segmentation, and counting.

  3. Reasoning-Guided Part-Level Visual Grounding via Reinforcement Learning

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A 4B MLLM trained with object-first hierarchical prompting and stage-wise part-aware GRPO rewards outperforms 7B grounding LLMs and SAM3 on part-grounding benchmarks.

  4. DGSeg: Dynamic Gating of Semantic-Spatial Guided Predictions for Reasoning Segmentation

    cs.CV 2026-07 accept novelty 6.0 of 10

    Separate semantic-spatial MLLM cues plus a quality-supervised dynamic gate yield 69.6/67.3 gIoU on ReasonSeg val/test, beating strong MLLM+SAM baselines.

Reference graph

Works this paper leans on

78 extracted references · 53 canonical work pages · cited by 4 Pith papers

  1. [1]

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023

  2. [2]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y . Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y . Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

  3. [3]

    K. Chen, Z. Zhang, W. Zeng, R. Zhang, F. Zhu, and R. Zhao. Shikra: Unleashing multimodal llm’s referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023

  4. [4]

    L. Chen, L. Li, H. Zhao, Y . Song, and Vinci. R1-v: Reinforcing super generalization ability in vision-language models with less than $3. https://github.com/Deep-Agent/R1-V , 2025. Accessed: 2025-02-02

  5. [5]

    Cheng, I

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar. Masked-attention mask trans- former for universal image segmentation. CVPR, 2022

  6. [6]

    Cheng, A

    B. Cheng, A. Schwing, and A. Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 2021

  7. [7]

    R. Cong, Q. Qin, C. Zhang, Q. Jiang, S. Wang, Y . Zhao, and S. Kwong. A weakly supervised learning framework for salient object detection via hybrid labels. TCSVT, 2022

  8. [8]

    T. Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023

Show all 78 references
  1. [9]

    Fan, M.-M

    D.-P. Fan, M.-M. Cheng, Y . Liu, T. Li, and A. Borji. Structure-measure: A new way to evaluate foreground maps. In ICCV, 2017

  2. [10]

    D.-P. Fan, C. Gong, Y . Cao, B. Ren, M.-M. Cheng, and A. Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018

  3. [11]

    Fan, G.-P

    D.-P. Fan, G.-P. Ji, G. Sun, M.-M. Cheng, J. Shen, and L. Shao. Camouflaged object detection. In CVPR, 2020

  4. [12]

    C. Fu, P. Chen, Y . Shen, Y . Qin, M. Zhang, X. Lin, J. Yang, X. Zheng, K. Li, X. Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023

  5. [13]

    Gemini 2.5 technical report

    Google DeepMind. Gemini 2.5 technical report. https://storage.googleapis.com/ deepmind-media/gemini/gemini_v2_5_report.pdf, 2025. Accessed: 2025-06-18

  6. [14]

    J. Hu, Z. Cheng, and S. Gong. Int: Instance-specific negative mining for task-generic prompt- able segmentation. arXiv preprint arXiv:2501.18753, 2025

  7. [15]

    J. Hu, J. Lin, S. Gong, and W. Cai. Relax image-specific prompt requirement in sam: A single generic prompt for segmenting camouflaged objects. In AAAI, 2024

  8. [16]

    J. Hu, J. Lin, J. Yan, and S. Gong. Leveraging hallucinations to reduce manual prompt depen- dency in promptable segmentation. arXiv preprint arXiv:2408.15205, 2024

  9. [17]

    Huang, H

    Z. Huang, H. Dai, T.-Z. Xiang, S. Wang, H.-X. Chen, J. Qin, and H. Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In CVPR, 2023. 10

  10. [18]

    Kazemzadeh, V

    S. Kazemzadeh, V . Ordonez, M. Matten, and T. Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, 2014

  11. [19]

    Kembhavi, M

    A. Kembhavi, M. Salvato, E. Kolve, M. Seo, H. Hajishirzi, and A. Farhadi. A diagram is worth a dozen images. In ECCV, 2016

  12. [20]

    Kirillov, K

    A. Kirillov, K. He, R. Girshick, C. Rother, and P. Dollár. Panoptic segmentation. In CVPR, 2019

  13. [21]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, et al. Segment anything. InICCV, 2023

  14. [22]

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu, J. Gonzalez, H. Zhang, and I. Stoica. Efficient memory management for large language model serving with pagedattention. InSOSP, 2023

  15. [23]

    X. Lai, Z. Tian, Y . Chen, Y . Li, Y . Yuan, S. Liu, and J. Jia. Lisa: Reasoning segmentation via large language model. In CVPR, 2024

  16. [24]

    T.-N. Le, T. V . Nguyen, Z. Nie, M.-T. Tran, and A. Sugimoto. Anabranch network for camou- flaged object segmentation. CVIU, 2019

  17. [25]

    Li and Y

    G. Li and Y . Yu. Visual saliency based on multiscale deep features. In CVPR, 2015

  18. [26]

    J. Li, D. Li, S. Savarese, and S. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023

  19. [27]

    J. Li, D. Li, C. Xiong, and S. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888–12900. PMLR, 2022

  20. [28]

    Y . Li, Y . Du, K. Zhou, J. Wang, W. X. Zhao, and J.-R. Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023

  21. [29]

    Liang, B

    F. Liang, B. Wu, X. Dai, K. Li, Y . Zhao, H. Zhang, P. Zhang, P. Vajda, and D. Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023

  22. [30]

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014

  23. [31]

    C. Liu, H. Ding, and X. Jiang. Gres: Generalized referring expression segmentation. In CVPR, 2023

  24. [32]

    H. Liu, C. Li, Q. Wu, and Y . J. Lee. Visual instruction tuning. NeurIPS, 2023

  25. [33]

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In ECCV, 2024

  26. [34]

    W. Liu, X. Shen, C.-M. Pun, and X. Cun. Explicit visual prompting for low-level structure segmentations. In CVPR, 2023

  27. [35]

    W. Liu, X. Shen, C.-M. Pun, and X. Cun. Explicit visual prompting for universal foreground segmentations. arXiv preprint arXiv:2305.18476, 2023

  28. [36]

    Y . Liu, H. Duan, Y . Zhang, B. Li, S. Zhang, W. Zhao, Y . Yuan, J. Wang, C. He, Z. Liu, et al. Mmbench: Is your multi-modal model an all-around player? In ECCV, 2024

  29. [37]

    Y . Liu, Y . Zhang, Z. Wang, F. Yang, F. Qiu, S. Coleman, and D. Kerr. A novel seminar learning framework for weakly supervised salient object detection. EAAI, 2023

  30. [38]

    M. Ma, C. Xia, C. Xie, X. Chen, and J. Li. Receptive field broadening and boosting for salient object detection. arXiv preprint arXiv:2110.07859, 2021

  31. [39]

    Mei, G.-P

    H. Mei, G.-P. Ji, Z. Wei, X. Yang, X. Wei, and D.-P. Fan. Camouflaged object segmentation with distraction mining. In CVPR, 2021. 11

  32. [40]

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep rein- forcement learning. nature, 2015

  33. [41]

    Gpt-4 technical report, 2023

    OpenAI. Gpt-4 technical report, 2023

  34. [42]

    Ouyang, J

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022

  35. [43]

    Y . Pang, X. Zhao, T.-Z. Xiang, L. Zhang, and H. Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, 2022

  36. [44]

    Y . Pang, X. Zhao, T.-Z. Xiang, L. Zhang, and H. Lu. Zoomnext: A unified collaborative pyramid network for camouflaged object detection. TPAMI, 2024

  37. [45]

    Y . Piao, W. Wu, M. Zhang, Y . Jiang, and H. Lu. Noise-sensitive adversarial learning for weakly supervised salient object detection. TMM, 2022

  38. [46]

    B. A. Plummer, L. Wang, C. M. Cervantes, J. C. Caicedo, J. Hockenmaier, and S. Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015

  39. [47]

    X. Qin, H. Dai, X. Hu, D.-P. Fan, L. Shao, and L. V . Gool. Highly accurate dichotomous image segmentation. In ECCV, 2022

  40. [48]

    Rafailov, A

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2023

  41. [49]

    Rasheed, M

    H. Rasheed, M. Maaz, S. Shaji, A. Shaker, S. Khan, H. Cholakkal, R. M. Anwer, E. Xing, M.-H. Yang, and F. S. Khan. Glamm: Pixel grounding large multimodal model. In CVPR, pages 13009–13018, 2024

  42. [50]

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Dol- lár, and C. Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408...

  43. [51]

    T. Ren, S. Liu, A. Zeng, J. Lin, K. Li, H. Cao, J. Chen, X. Huang, Y . Chen, F. Yan, Z. Zeng, H. Zhang, F. Li, J. Yang, H. Li, Q. Jiang, and L. Zhang. Grounded sam: Assembling open- world models for diverse visual tasks, 2024

  44. [52]

    Z. Ren, Z. Huang, Y . Wei, Y . Zhao, D. Fu, J. Feng, and X. Jin. Pixellm: Pixel reasoning with large multimodal model. In CVPR, pages 26374–26383, 2024

  45. [53]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

  46. [54]

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

  47. [55]

    H. Shen, P. Liu, J. Li, C. Fang, Y . Ma, J. Liao, Q. Shen, Z. Zhang, K. Zhao, Q. Zhang, et al. Vlm-r1: A stable and generalizable r1-style large vision-language model. arXiv preprint arXiv:2504.07615, 2025

  48. [56]

    J. Shi, Q. Yan, L. Xu, and J. Jia. Hierarchical image saliency detection on extended cssd. TPAMI, 2015

  49. [57]

    R. S. Sutton, A. G. Barto, et al. Reinforcement learning: An introduction . MIT press Cam- bridge, 1998

  50. [58]

    J. Wang, Z. Tian, X. Wang, X. Zhang, W. Huang, Z. Wu, and Y .-G. Jiang. Simplear: Pushing the frontier of autoregressive visual generation through pretraining, sft, and rl. arXiv preprint arXiv:2504.11455, 2025. 12

  51. [59]

    L. Wang, H. Lu, Y . Wang, M. Feng, D. Wang, B. Yin, and X. Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017

  52. [60]

    P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y . Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, and J. Lin. Qwen2-vl: En- hancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409....

  53. [61]

    C. Yang, L. Zhang, H. Lu, X. Ruan, and M.-H. Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013

  54. [62]

    Q. Ye, H. Xu, G. Xu, J. Ye, M. Yan, Y . Zhou, J. Wang, A. Hu, P. Shi, Y . Shi, et al. mplug- owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023

  55. [63]

    Z. You, L. Kong, L. Meng, and Z. Wu. FOCUS: Towards universal foreground segmentation. In AAAI, 2025

  56. [64]

    Z. You, J. Wang, L. Kong, B. He, and Z. Wu. Pix2cap-coco: Advancing visual comprehension via pixel-level captioning. arXiv preprint arXiv:2501.13893, 2025

  57. [65]

    L. Yu, P. Poirson, S. Yang, A. C. Berg, and T. L. Berg. Modeling context in referring expres- sions. In ECCV, 2016

  58. [66]

    H. Yuan, X. Li, T. Zhang, Z. Huang, S. Xu, S. Ji, Y . Tong, L. Qi, J. Feng, and M.-H. Yang. Sa2va: Marrying sam2 with llava for dense grounded understanding of images and videos. arXiv preprint arXiv:2501.04001, 2025

  59. [67]

    Y . Yuan, W. Liu, P. Gao, Q. Dai, and J. Qin. Unified unsupervised salient object detection via knowledge transfer. arXiv preprint arXiv:2404.14759, 2024

  60. [68]

    Y . K. Yun and W. Lin. Towards a complete and detail-preserved salient object detection.TMM, 2023

  61. [69]

    Zhang, P

    S. Zhang, P. Sun, S. Chen, M. Xiao, W. Shao, W. Zhang, Y . Liu, K. Chen, and P. Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. In ECCV, 2025

  62. [70]

    Zhang, X

    T. Zhang, X. Li, H. Fei, H. Yuan, S. Wu, S. Ji, C. C. Loy, and S. Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. NeurIPS, 2024

  63. [71]

    Zhang, X

    T. Zhang, X. Li, Z. Huang, Y . Li, W. Lei, X. Deng, S. Chen, S. Ji, and J. Feng. Pixel-sail: Single transformer for pixel-grounded understanding. arXiv preprint arXiv:2504.10465, 2025

  64. [72]

    Zheng, D

    P. Zheng, D. Gao, D.-P. Fan, L. Liu, J. Laaksonen, W. Ouyang, and N. Sebe. Bilateral reference for high-resolution dichotomous image segmentation. arXiv preprint arXiv:2401.03407, 2024

  65. [73]

    B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso, and A. Torralba. Scene parsing through ade20k dataset. In CVPR, 2017

  66. [74]

    H. Zhou, P. Chen, L. Yang, X. Xie, and J. Lai. Activation to saliency: Forming high-quality labels for unsupervised salient object detection. TCSVT, 2022

  67. [75]

    H. Zhou, B. Qiao, L. Yang, J. Lai, and X. Xie. Texture-guided saliency distilling for unsuper- vised salient object detection. In CVPR, 2023

  68. [76]

    H. Zhu, P. Li, H. Xie, X. Yan, D. Liang, D. Chen, M. Wei, and J. Qin. I can find you! boundary- guided separated attention network for camouflaged object detection. In AAAI, 2022

  69. [77]

    Zou, Z.-Y

    X. Zou, Z.-Y . Dou, J. Yang, Z. Gan, L. Li, C. Li, X. Dai, H. Behl, J. Wang, L. Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, 2023

  70. [78]

    X. Zou, J. Yang, H. Zhang, F. Li, L. Li, J. Wang, L. Wang, J. Gao, and Y . J. Lee. Segment everything everywhere all at once. NeurIPS, 2023. 13

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.