pith. sign in

arxiv: 2607.00544 · v1 · pith:SFTMJJTYnew · submitted 2026-07-01 · 💻 cs.CV

GEAR-Seg: A Grounded Explainable Agent for Reasoning Segmentation and Data Engine

Yanan Wang , Wen Li , Yibin Ying , Zhenghao Fei This is my paper

Pith reviewed 2026-07-02 14:31 UTC · model grok-4.3

classification 💻 cs.CV
keywords reasoning segmentationexplainable AILLM deductiondata enginezero-shot inferencereferring segmentationmultimodal reasoningsynthetic data
0
0 comments X

The pith

GEAR-Seg decouples segmentation, text description, and LLM deduction to turn implicit reasoning into an explicit logic chain.

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

Reasoning segmentation localizes objects from complex implicit queries, but end-to-end models hide the steps inside an opaque box that limits both understanding and scaling. GEAR-Seg splits the task into three separate stages: class-agnostic region finding, conversion of those regions into dense attribute text, and LLM deduction over the text. The separation produces a trackable chain of steps instead of a black-box answer. The same pipeline doubles as a data engine that automatically labels over 38,000 images with 656,000 QA-mask pairs, creating the GEAR-131K benchmark. Distilled lightweight models trained only on this synthetic data reach performance close to models trained on expensive human labels.

Core claim

By decoupling class-agnostic segmentation, semantic description, and Large Language Model deduction, GEAR-Seg converts implicit visual reasoning into an explicit, trackable logic chain. As a zero-shot framework it matches competitive performance on reasoning and referring segmentation benchmarks. The same architecture functions as a scalable data engine that produces the GEAR-131K benchmark containing more than 38k images and 656k QA-mask pairs organized under a manipulation-oriented taxonomy. Distillation experiments show that models trained solely on the automatically generated data approach the accuracy of models trained on human-annotated data.

What carries the argument

Three-stage decoupled pipeline that first extracts class-agnostic regions, then renders each region as attribute-rich text, then applies LLM deduction on the resulting text descriptions.

Load-bearing premise

Converting visual regions into dense attribute-rich text descriptions preserves all information needed for accurate LLM deduction on complex implicit queries without introducing critical omissions or hallucinations.

What would settle it

On a held-out set of complex implicit queries, measure whether GEAR-Seg's LLM deductions systematically miss targets that a direct end-to-end model correctly segments; a large gap would indicate information loss in the text step.

Figures

Figures reproduced from arXiv: 2607.00544 by Wen Li, Yanan Wang, Yibin Ying, Zhenghao Fei.

Figure 1
Figure 1. Figure 1: Overview of GEAR-Seg’s multifaceted capabilities. Serving as both a zero-shot inference agent and a scalable data engine, it explicitly translates pixels into text to seamlessly support complex reasoning segmentation, dense referring segmentation, and fine-grained attribute grounding in long-tail domains. Despite rapid progress, current state-of-the-art (SOTA) architectures typi￾cally formulate reasoning s… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the GEAR-Seg framework. The agent explicitly decouples the reasoning segmentation task into class-agnostic perception (SAM 2), dense semantic description (DAM), and logic-driven abstraction (LLM), serving as both a zero-shot inference engine and a scalable data generator. effective paradigm. Instead of directly distilling model weights, this approach uti￾lizes a powerful agent as a teacher to a… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the GEAR-Seg data generation pipeline and operational modes. synthesizing these comprehensive modalities to autonomously generate a diverse set of annotations. For each image, the engine outputs a challenging base query, an explicit step-by-step logic chain, and the corresponding precise mask indices, thereby establishing a high-quality benchmark for reasoning segmentation. 4.2 Taxonomy of Reas… view at source ↗
Figure 4
Figure 4. Figure 4: Detailed statistics of the GEAR-131K benchmark. (a) Image distribution across source datasets. (b) Proportion of the five specialized reasoning categories. (c) Word cloud illustrating the semantic diversity of the targeted entities. (d) Comprehensive feature comparison against existing reasoning segmentation datasets. Mapillary [20], and ADE20K [37]. Our automated engine initially generated 162k raw propos… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of the GEAR-Seg agent. (a) Complex reasoning segmen￾tation on ReasonSeg and LLM-Seg40k. (b) Open-world auto-label extraction across diverse agricultural scenes, showcasing the zero-shot discovery of long-tail categories. (c) Fine-grained maturity grading, demonstrating precise attribute-based grounding under severe occlusion. Plug-and-Play Cognitive Flexibility. Unlike end-to-end black-… view at source ↗
Figure 6
Figure 6. Figure 6: Accuracy evaluation and typical failure modes of the GEAR-Seg agent, illus￾trating cascading errors in complex reasoning and attribute hallucination. making it a highly worthwhile tradeoff for offline dataset generation and complex multi-turn reasoning tasks. 5.4 Knowledge Distillation to End-to-End Models To fully unleash the potential of the massive datasets generated by our data engine, we conduct knowl… view at source ↗
Figure 7
Figure 7. Figure 7: Top: Representative examples of the 5-fold linguistic expansion in the GEAR￾131K dataset. Bottom: Additional dataset visualizations of GEAR-131K. other relevant semantic elements in the scene. Following this instruction, GEAR￾Seg independently analyzes the global visual context, identifies fine-grained cat￾egories, and assigns appropriate text labels to all instances. To ensure a rigorous and unbiased zero… view at source ↗
Figure 8
Figure 8. Figure 8: Examples of the auto-label extracting capability of GEAR-Seg across diverse MegaFruits datasets. By prompting the model to analyze the scene context, it auto￾matically extracts a set of semantic categories and assigns fine-grained labels to each detected instance, including long-tail objects often missed by human annotators. Seg accurately identifies and labels diseased leaf (fig. 8b), as well as structura… view at source ↗
read the original abstract

Reasoning segmentation requires localizing targets based on complex, implicit queries. Current end-to-end models typically entangle perception and deduction into an opaque black box, severely limiting interpretability and scalability. To address this, we propose GEAR-Seg (Grounded Explainable Agent for Reasoning Segmentation), an explicitly decoupled agent that shifts the paradigm by translating visual pixels into dense, attribute-rich text. By decoupling class-agnostic segmentation, semantic description, and Large Language Model (LLM) deduction, GEAR-Seg transforms implicit reasoning into an explicit, trackable logic chain. As a zero-shot inference framework, it achieves highly competitive performance across diverse reasoning and fine-grained referring segmentation benchmarks. Furthermore, GEAR-Seg inherently functions as a highly scalable data engine. Utilizing this engine, we construct GEAR-131K, a massive benchmark (over 38k images, 656k QA-mask pairs) introducing a multifaceted taxonomy tailored for complex real-world manipulation-oriented reasoning. Finally, distillation experiments demonstrate that lightweight models supervised exclusively by our automated pipeline closely match the upper-bound performance of costly human-annotated baselines.

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 / 1 minor

Summary. The paper proposes GEAR-Seg, a decoupled agent framework for reasoning segmentation that separates class-agnostic mask generation, VLM-based dense semantic description of regions, and LLM-based deduction to produce an explicit, trackable reasoning chain. It claims competitive zero-shot performance on reasoning and fine-grained referring segmentation benchmarks, positions the method as a scalable data engine to create the GEAR-131K benchmark (38k+ images, 656k QA-mask pairs with a manipulation-oriented taxonomy), and reports that lightweight models distilled from the automated pipeline match human-annotated upper bounds.

Significance. If the empirical claims hold with supporting evidence, the work would offer a concrete advance in interpretability for complex vision-language reasoning tasks and a practical route to large-scale automated dataset creation, reducing reliance on costly human annotations while maintaining performance.

major comments (2)
  1. [Abstract] Abstract: The central claims of 'highly competitive performance across diverse reasoning and fine-grained referring segmentation benchmarks' and that 'lightweight models supervised exclusively by our automated pipeline closely match the upper-bound performance of costly human-annotated baselines' are stated without any quantitative tables, benchmark scores, error bars, or ablation results; this absence prevents evaluation of the empirical soundness of the zero-shot and distillation results.
  2. [Section 3] Section 3 (method): The pipeline generates per-region captions via a vision-language model and concatenates them as input to the LLM for deduction; this step assumes the text descriptions preserve all spatial relations, occlusion details, texture gradients, and context needed for accurate deduction on implicit manipulation queries, yet no validation, failure-case analysis, or comparison against direct visual input is provided to test this assumption, which is load-bearing for the 'explicit, trackable logic chain' claim.
minor comments (1)
  1. [Abstract] Abstract: The dataset is described as 'GEAR-131K' with 'over 38k images, 656k QA-mask pairs'; the naming convention and exact scope of the 131K figure should be clarified relative to the reported counts.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for quantitative support in the abstract and validation of the text-based reasoning assumption. We address both points below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claims of 'highly competitive performance across diverse reasoning and fine-grained referring segmentation benchmarks' and that 'lightweight models supervised exclusively by our automated pipeline closely match the upper-bound performance of costly human-annotated baselines' are stated without any quantitative tables, benchmark scores, error bars, or ablation results; this absence prevents evaluation of the empirical soundness of the zero-shot and distillation results.

    Authors: We agree that the abstract would be strengthened by including key quantitative results. In the revised version, we will add specific benchmark scores (e.g., mIoU on reasoning segmentation tasks and comparison to baselines) with references to the main tables, while keeping the abstract concise. This directly addresses the concern about empirical soundness. revision: yes

  2. Referee: [Section 3] Section 3 (method): The pipeline generates per-region captions via a vision-language model and concatenates them as input to the LLM for deduction; this step assumes the text descriptions preserve all spatial relations, occlusion details, texture gradients, and context needed for accurate deduction on implicit manipulation queries, yet no validation, failure-case analysis, or comparison against direct visual input is provided to test this assumption, which is load-bearing for the 'explicit, trackable logic chain' claim.

    Authors: The referee correctly notes that the captioning step is central to the explicit chain. The current manuscript does not include a dedicated validation study or direct comparison to visual-input baselines. We will add a new subsection with quantitative comparison of LLM deduction accuracy using VLM captions versus direct image input, plus failure-case analysis on spatial/occlusion details. This will either support the assumption or clarify its limitations. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is empirical and self-contained against external benchmarks

full rationale

The paper describes GEAR-Seg as a zero-shot decoupled framework (class-agnostic masks + VLM captions + LLM deduction) evaluated on external reasoning/referring segmentation benchmarks, with the data engine used to generate new GEAR-131K data and distillation results compared to human baselines. No equations, fitted parameters, or predictions are presented that reduce reported performance or claims to the inputs by construction. No self-citation load-bearing steps or ansatz smuggling appear in the provided text. The central claims rest on empirical outcomes and the explicit decoupling architecture rather than self-referential definitions or renamings of known results.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the domain assumption that LLMs can reliably perform deduction from generated text descriptions; no free parameters or invented physical entities are mentioned. Full paper would be needed to audit any additional modeling choices.

axioms (1)
  • domain assumption Large language models can perform accurate deduction on complex implicit queries when given dense attribute-rich text descriptions of image regions.
    Central to the claim that the decoupled pipeline preserves reasoning capability.
invented entities (1)
  • GEAR-Seg agent no independent evidence
    purpose: Explicitly decoupled pipeline for grounded explainable reasoning segmentation
    New system architecture proposed by the authors.

pith-pipeline@v0.9.1-grok · 5733 in / 1409 out tokens · 26494 ms · 2026-07-02T14:31:05.121276+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

38 extracted references · 29 canonical work pages · 3 internal anchors

  1. [1]

    , year = 2025, journal =

    Acharya, D.B., Kuppan, K., Divya, B.: Agentic AI: Autonomous intelligence for complex goals—a comprehensive survey. IEEE Access13, 18912–18936 (2025). https://doi.org/10.1109/ACCESS.2025.3532853

    work page doi:10.1109/access.2025.3532853 2025

  2. [2]

    In: IEEE Conf

    Chen, R., Li, C., Wu, Q., Zhong, Y.Z., Han, P., Li, W., Wei, Y., Zhao, Y.: LLM- Seg: Bridging image segmentation and large language model reasoning. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 1765–1774 (2024).https://doi. org/10.1109/CVPRW63382.2024.00183

    work page doi:10.1109/cvprw63382.2024.00183 2024

  3. [3]

    Capsfusion: Rethinking image-text data at scale

    Chen, X., Hu, J., Chen, Z., Li, Y., Darrell, T., Yu, F., Gao, J.: LISA: Reasoning segmentation via large language models. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 9579–9589 (2024).https://doi.org/10.1109/CVPR52733.2024.00915

    work page doi:10.1109/cvpr52733.2024.00915 2024

  4. [4]

    Chen, Y.C., Li, W.H., Sun, C., Wang, Y.C.F., Chen, C.S.: SAM4MLLM: En- hance multi-modal large language model for referring expression segmentation. In: Eur. Conf. Comput. Vis. pp. 323–340 (2024).https://doi.org/10.1007/ 978-3-031-73004-7_19 16 Y. Wang et al

    2024

  5. [5]

    Capsfusion: Rethinking image-text data at scale

    Cheng, T., Song, L., Ge, Y., Liu, W., Wang, X., Shan, Y.: YOLO-World: Real-time open-vocabulary object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16901–16911 (2024).https://doi.org/10.1109/CVPR52733.2024.01599

    work page doi:10.1109/cvpr52733.2024.01599 2024

  6. [6]

    Ding, H., Liu, C., Wang, S., Jiang, X.: Vision-language transformer and query generation for referring segmentation. In: Int. Conf. Comput. Vis. pp. 16301–16310 (2021).https://doi.org/10.1109/ICCV48922.2021.01601

    work page doi:10.1109/iccv48922.2021.01601 2021

  7. [7]

    Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes (VOC) challenge. Int. J. Comput. Vis.88(2), 303– 338 (2010).https://doi.org/10.1007/s11263-009-0275-4

    work page doi:10.1007/s11263-009-0275-4 2010

  8. [8]

    The Llama 3 Herd of Models

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Let- man, A., Mathur, A., Schelten, A., Vaughan, A., et al.: The Llama 3 herd of mod- els. arXiv preprint arXiv:2407.21783 (2024).https://doi.org/10.48550/arXiv. 2407.21783

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2024

  9. [9]

    In: IEEE Conf

    Gupta, A., Doll´ ar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5356–5364 (2019). https://doi.org/10.1109/CVPR.2019.00550

    work page doi:10.1109/cvpr.2019.00550 2019

  10. [10]

    Hu, R., Rohrbach, M., Darrell, T.: Segmentation from natural language expres- sions. In: Eur. Conf. Comput. Vis. pp. 108–124 (2016).https://doi.org/10.1007/ 978-3-319-46448-0_7

    2016

  11. [11]

    Jang, D., Cho, Y., Lee, S., Kim, T., Kim, D.: MMR: A large-scale benchmark dataset for multi-target and multi-granularity reasoning segmentation. In: Int. Conf. Learn. Represent. (2025),https://openreview.net/forum?id=mzL19kKE3r

    2025

  12. [12]

    Kirillov, A., Girshick, R.M., Doll´ ar, P., Mahajan, D.R., et al.: Segment anything. In: Int. Conf. Comput. Vis. pp. 4015–4026 (2023).https://doi.org/10.1109/ ICCV51070.2023.00371

    work page arXiv 2023

  13. [13]

    In: Lecture Notes in Networks and Systems

    Kozlov, A., Lazarevich, I., Shamporov, V., Lyalyushkin, N., Gorbachev, Y.: Neural network compression framework for fast model inference. In: Lecture Notes in Networks and Systems. vol. 285, pp. 240–253 (2021).https://doi.org/10.1007/ 978-3-030-80129-8_17

    2021

  14. [14]

    In: IEEE Conf

    Li, Y., Chen, C., Dai, X., Chen, H.: Overcoming classifier imbalance for long- tail object detection with balanced group softmax. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10988–10997 (2020).https://doi.org/10.1109/CVPR42600. 2020.01100

    work page doi:10.1109/cvpr42600 2020

  15. [15]

    Lian, L., Ding, Y., Ge, Y., Cui, Y., Yala, A., Darrell, T.: DAM: Describe anything model for detailed localized image and video captioning. In: Int. Conf. Comput. Vis. pp. 21766–21777 (2025)

    2025

  16. [16]

    Capsfusion: Rethinking image-text data at scale

    Liang, Y., Li, C., Zhang, D., Yang, Z., Wang, B., Mei, T.: CogAgent: A visual language model for GUI agents. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14281–14290 (2024).https://doi.org/10.1109/CVPR52733.2024.01354

    work page doi:10.1109/cvpr52733.2024.01354 2024

  17. [17]

    In: Conference on Computer Vision and Pattern Recognition (CVPR)

    Liu, C., Ding, H., Jiang, X.: GRES: Generalized referring expression segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 23592–23601 (2023).https: //doi.org/10.1109/CVPR52729.2023.02259

    work page doi:10.1109/cvpr52729.2023.02259 2023

  18. [18]

    In: IEEE Conf

    Liu, Y., Zhang, J., Han, J., Yang, Y., Li, C., Gao, J.: LAVT: Language-aware vision transformer for referring image segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 18134–18144 (2022).https://doi.org/10.1109/CVPR52688. 2022.01762

    work page doi:10.1109/cvpr52688 2022

  19. [19]

    Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Doso- vitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., Houlsby, N.: Simple Open-Vocabulary object detection. In: Eur. Conf. Comput. Vis. pp. 728–755 (2022).https://doi.org/10.1007/ 978-3-031-20080-9_42 GEAR-Seg: A Grounded Explainable Agent fo...

    2022

  20. [20]

    Neuhold, G., Ollmann, T., Rota Bulo, S., Kontschieder, P.: The Mapillary Vistas dataset for semantic understanding of street scenes. In: Int. Conf. Comput. Vis. pp. 5122–5130 (2017).https://doi.org/10.1109/ICCV.2017.534

    work page doi:10.1109/iccv.2017.534 2017

  21. [21]

    P´ erez-Borrero, I., Mar´ ın-Santos, D., Geg´ undez-Arias, M.E., Cort´ es-Ancos, E.: A fast and accurate deep learning method for strawberry instance segmenta- tion. Comput. Electron. Agric.178, 105736 (2020).https://doi.org/10.1016/ j.compag.2020.105736

    work page arXiv 2020

  22. [22]

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., R¨ adle, R., Rolland, C., Gustafson, L., et al.: SAM 2: Segment anything in images and videos. In: Int. Conf. Learn. Represent. (2024).https://doi.org/10.48550/arXiv.2408. 00714

    work page doi:10.48550/arxiv.2408 2024

  23. [23]

    Reimers, N., Gurevych, I.: Sentence-BERT: Sentence embeddings using Siamese BERT-networks. pp. 3982–3992 (2019).https://doi.org/10.18653/v1/D19-1410

    work page doi:10.18653/v1/d19-1410 2019

  24. [24]

    Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks

    Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.: Grounded SAM: Assembling open-world models for diverse visual tasks. In: arXiv preprint arXiv:2401.14159 (2024).https://doi.org/10.48550/ arXiv.2401.14159

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  25. [25]

    Capsfusion: Rethinking image-text data at scale

    Ren, Z., Huang, Z., Wei, Y., Zhao, Y., Fu, D., Feng, J., Jin, X.: PixelLM: Pixel rea- soning with large multimodal model. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 26374–26383 (2024).https://doi.org/10.1109/CVPR52733.2024.02491

    work page doi:10.1109/cvpr52733.2024.02491 2024

  26. [26]

    In: IEEE Conf

    Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Gen- eralized Intersection Over Union: A metric and a loss for bounding box regres- sion. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 658–666 (2019).https: //doi.org/10.1109/CVPR.2019.00075

    work page doi:10.1109/cvpr.2019.00075 2019

  27. [27]

    Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomed- ical image segmentation. In: Med. Image Comput. Comput.-Assist. Intervent. pp. 234–241 (2015).https://doi.org/10.1007/978-3-319-24574-4_28

    work page doi:10.1007/978-3-319-24574-4_28 2015

  28. [28]

    Sachdeva, N., Dhaliwal, M., Wu, C.J., McAuley, J.: Infinite Recommendation Net- works: A data-centric approach. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 31292–31305 (2022)

    2022

  29. [29]

    M., S.: YOLOv8: A novel object detection algorithm with enhanced performance and robustness

    Varghese, R., S. M., S.: YOLOv8: A novel object detection algorithm with enhanced performance and robustness. In: International Conference on Artificial Intelligence and Data Sciences. pp. 1–6 (2024).https://doi.org/10.1109/ADICS58448.2024. 10533619

    work page doi:10.1109/adics58448.2024 2024

  30. [30]

    Pattern Recognition174, 112799 (2026)

    Wang, Y., Fei, Z., Li, R., Ying, Y.: Learn from foundation model: Fruit detec- tion model without manual annotation. Pattern Recognition174, 112799 (2026). https://doi.org/10.1016/j.patcog.2025.112799

    work page doi:10.1016/j.patcog.2025.112799 2026

  31. [31]

    Ego4d: Around the world in 3, 000 hours of egocentric video

    Wang, Z., Lu, Y., Li, Q., Tao, X., Guo, Y., Gong, M., Liu, T.: CRIS: CLIP-driven referring image segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11676–11685 (2022).https://doi.org/10.1109/CVPR52688.2022.01139

    work page doi:10.1109/cvpr52688.2022.01139 2022

  32. [32]

    Qwen3 Technical Report

    Yang, A., Li, A., Yang, B., et al.: Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025).https://doi.org/10.48550/arXiv.2505.09388

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.09388 2025

  33. [33]

    In: IEEE Conf

    Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: MAttNet: Modular attention network for referring expression comprehension. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1307–1315 (2018)

    2018

  34. [34]

    Agentic AI: A conceptual taxonomy, applica- tions and challenges

    Zhang, L., et al.: AI agents vs. Agentic AI: A conceptual taxonomy, applica- tions and challenges. Information Fusion122, 103599 (2025).https://doi.org/ 10.1016/j.inffus.2025.103599

    work page doi:10.1016/j.inffus.2025.103599 2025

  35. [35]

    Lambourne, Karl D

    Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H.S., et al.: Rethinking semantic segmentation from a sequence-to-sequence 18 Y. Wang et al. perspective with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6877–6886 (2021).https://doi.org/10.1109/CVPR46437.2021.00681

    work page doi:10.1109/cvpr46437.2021.00681 2021

  36. [36]

    In: AAAI

    Zheng, Z., Wang, P., Liu, W., Li, J., Ye, R., Ren, D.: Distance-IoU loss: Faster and better learning for bounding box regression. In: AAAI. pp. 12993–13000 (2020). https://doi.org/10.1609/aaai.v34i07.6999

    work page doi:10.1609/aaai.v34i07.6999 2020

  37. [37]

    5122–5130

    Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A.: Scene parsing through ADE20K dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5122– 5130 (2017).https://doi.org/10.1109/CVPR.2017.544

    work page doi:10.1109/cvpr.2017.544 2017

  38. [38]

    Zhu, L., Chen, T., Xu, Q., Liu, X., Ji, D., Wu, H., Soh, D.W., Liu, J.: Popen: Preference-based optimization and ensemble for LVLM-based reasoning segmen- tation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) GEAR-Seg: A Grounded Explainable Agent for Reasoning Segmentation 19 A Supplementary Material ...

    2025