Pith. sign in

REVIEW 3 major objections 5 minor 39 references

This paper shows that referring video object segmentation can be solved without any training: a compact language model, reading object trajectories as text, outperforms trained specialists on three benchmarks, reaching J&F of 72.1 on Ref-Yo

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A training-free pipeline parses a video query with an LLM, tracks candidates with SAM2, then uses coarse motion plus conditional pose reasoning to segment the referred object.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A serious training-free RVOS result with a real appendix contradiction: the motion-reasoning prompt forces a single candidate ID while the method requires a filtered subset, so multi-target cases are not reproducible as written. the 3 major comments →

arxiv 2509.05751 v1 pith:OE7B57UL submitted 2025-09-06 cs.CV cs.AI

Unleashing Hierarchical Reasoning: An LLM-Driven Framework for Training-Free Referring Video Object Segmentation

Bingrui Zhao , Lin Yuanbo Wu , Xiangtian Fan , Deyin Liu , Lu Zhang , Ruyi He , Jialie Shen , Ximing Li This is my paper
classification cs.CV cs.AI
keywords referring video object segmentationtraining-freelarge language modelshierarchical reasoningmotion reasoningpose verificationzero-shot segmentationvideo object grounding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

PARSE-VOS tries to prove that referring video object segmentation can be solved without training a specialist model: a compact, off-the-shelf language model reasons over textualized object trajectories and outperforms trained competitors. The paper reports the best published J&F scores on three benchmarks—72.1 on Ref-YouTube-VOS, 75.5 on Ref-DAVIS17, and 52.4 on MeViS—while freezing all components. The argument's core is a hierarchy: parse the query into entity, motion, and pose commands; detect and track every candidate; then use the LLM to filter by motion with camera and occlusion priors, and only if ambiguity remains, verify fine-grained pose with CLIP. A sympathetic reading is that the semantic gap in RVOS is a reasoning gap, not a representation gap, and that explicit contextual priors let a general LLM close it. Why it matters: if the claim holds, RVOS systems can be assembled from interchangeable off-the-shelf parts and updated without retraining.

Core claim

On the paper's own terms, the discovery is that a training-free pipeline can reach state-of-the-art referring video object segmentation by converting the task into hierarchical reasoning. Candidate object trajectories are generated by open-vocabulary detection plus video segmentation, serialized into timestamped bounding-box strings, and given to Llama-3-8B together with an affine camera-motion estimate and an occlusion ordering computed from pixel counts. The LLM first performs coarse motion reasoning to filter candidates; only if more than K candidates survive does a conditional fine-grained stage crop three maximally separated keyframes and match them via CLIP to the pose/attribute query.

What carries the argument

The central mechanism is the hierarchical identification module (M3) operating on serialized trajectories. It converts each candidate's bounding-box sequence into text, adds two contextual priors—an affine camera-motion model fitted by sparse optical flow (Eq. 1) and a pixel-cardinality occlusion ordering—and prompts the LLM to reason coarse-to-fine. Coarse motion reasoning filters candidates using the motion command; the conditional pose-verification stage, activated only when ambiguity remains (|C′| > K), crops the three most visually separated keyframes and uses CLIP cosine similarity (Eq. 2) to pick the target. The priors are what allow the LLM to distinguish apparent motion caused by th

Load-bearing premise

The framework assumes that serialized bounding boxes, an affine camera-motion estimate, and a pixel-count occlusion ranking give an 8-billion-parameter LLM enough information to infer each object's intrinsic motion and select the right target without any training.

What would settle it

Synthesize videos where two candidates share identical intrinsic motion but differ only in a visual attribute invisible to bounding boxes (e.g., color or texture) and run the pipeline: if the conditional CLIP stage cannot resolve them while human annotators can, the fallback is insufficient. Also, on real videos with heavy camera motion, ablate the camera-motion paragraph from the LLM prompt; if the J&F drop is negligible, the affine model in Eq. (1) is not the load-bearing prior the paper claims.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Performance on RVOS benchmarks can be improved by swapping in better base perception models (detector, segmenter, CLIP) with no retraining, since the pipeline is modular and training-free.
  • The coarse-to-fine design reduces inference cost: motion filtering alone resolves many queries, and pose verification is triggered only for genuinely ambiguous cases.
  • The success of an 8B LLM over larger LLM-based systems suggests that structured contextual priors and staged reasoning can compensate for model scale.
  • The framework reframes RVOS from a representation-alignment problem to a reasoning problem, so the semantic-gap failures of holistic fusion become addressable by prompt design rather than dataset engineering.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because all components are frozen, the framework is a natural benchmark for perception-model progress: gains in open-vocabulary detection or video segmentation should transfer directly to RVOS without retraining.
  • The same parse-and-reason hierarchy may transfer to related grounded-video tasks such as referring video tracking, spatio-temporal grounding, or language-conditioned video editing, which also hinge on combining motion and attribute cues.
  • One testable extension is to enrich the LLM's input beyond bounding boxes—for example, feeding mask silhouettes or part-level states—to see whether fine-grained pose verification becomes more robust than the current CLIP crop averaging.
  • The camera-motion prior is an affine fit to sparse optical flow; comparing it with homography or depth-aware motion models on the same prompts would clarify how much of the gain comes from motion-model fidelity versus the LLM's reasoning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes PARSE-VOS, a training-free RVOS pipeline consisting of three modules: (M1) LLM-based decomposition of a referring expression into structured semantic commands; (M2) spatio-temporal candidate grounding, in which GroundingDINO+SAM2 detect and segment instances on sparse keyframes and a predictive association criterion links them into trajectories; and (M3) hierarchical target identification, in which an LLM performs coarse motion reasoning over serialized trajectories with camera-motion and occlusion priors and, conditionally, CLIP performs fine-grained pose verification. The paper reports J&F of 72.1 on Ref-YouTube-VOS, 75.5 on Ref-DAVIS17, and 52.4 on MeViS, claiming state-of-the-art performance on all three benchmarks.

Significance. If the reported results hold, the paper is significant: it would show that a compact 8B LLM used as a zero-shot spatio-temporal reasoner, combined entirely with off-the-shelf perception models, can outperform trained RVOS specialists, including larger LLM-based systems. The paper's strengths are that it is training-free, uses external public benchmarks, ships complete prompts and a reproducibility checklist, and does not rely on fitted equations. However, the central M3 module has a structural prompt inconsistency, the multi-target output protocol is unspecified, and the SOTA claim rests on single-run stochastic comparisons with hyperparameters tuned on the same validation set used for the headline MeViS number. These issues are load-bearing and need to be resolved before the claims can be accepted.

major comments (3)
  1. [§3.3 (M3) and Appendix Fig. 6(b)] The described cascade and the published prompt are incompatible. M3 states that coarse motion reasoning narrows C down to a subset C′ and that fine-grained pose verification is conditionally activated only when |C′|>K, with Eq. (2) then selecting from C′. However, the prompt in Fig. 6(b) instructs the LLM: “you must provide only the ID of the single best candidate object as your final answer.” This forces a single candidate, so C′ cannot contain multiple candidates, the |C′|>K branch is unreachable, and FPV cannot be triggered. It also cannot produce K>1 outputs for expressions such as “The two sea turtles...” in Fig. 8(b). If the deployed system uses a different prompt, the appendix does not report it; as written, the central reasoning module is unreproducible.
  2. [§3.3, Eq. (2), Fig. 8(b)] Even after correcting the prompt, no selection rule for K>1 is specified. Eq. (2) solves argmax over C′, returning a single target, and the pipeline diagram does not state how K targets are selected when |C′|>K and the parsed cardinality is K>1. MeViS contains multi-target referring expressions (e.g., Fig. 8(b)), so this is not a corner case. The paper must specify a top-K rule or an iterative removal procedure, and how the conditional FPV stage handles multiple remaining candidates.
  3. [Implementation Details, Appendix B Tables 4–5, Table 1] The SOTA claim is supported by insufficient statistical evidence. The LLM is run with temperature 0.7 and top_p 0.95 and no seed is reported; no number of runs or variance is given, despite reproducibility-checklist items 4.7 and 4.10 being marked “yes.” In addition, τ=15 and the verification window of 3 are selected by ablations on the MeViS validation set (Tables 4–5) and then used to report MeViS SOTA in Table 1. Since the margins over the nearest comparators are only 1.1–1.5 J&F points and the LLM is stochastic, the reported advantage could be within run-to-run noise or validation-set tuning artifacts. The authors should report multiple seeds/runs with confidence intervals and, ideally, separate hyperparameter selection from the reported evaluation.
minor comments (5)
  1. [Datasets, Table 1, abstract] Dataset names are inconsistent: “Ref-DA VIS17”, “Ref-Davis17”, “MeVis” and “MeViS” are used interchangeably, and “Vidoe-Lisa” appears in Table 1. Please standardize.
  2. [Eq. (1), §3.3] The text calls this a “robust sparse optical flow” estimate, but Eq. (1) is a plain least-squares photometric error minimization. Please clarify whether robust estimation (e.g., RANSAC or an M-estimator) is used, and whether the affine model is applied forward or backward. Also, Fig. 4's “camera zooms in to the left/right” is physically unclear; this likely means translation or pan.
  3. [Trajectory Formation and Temporal Association, Implementation Details] The association thresholds θ_iou and θ_dist are given only as “e.g., 0.6” and “e.g., 50 pixels” in the main text, and exact final values are not listed in Implementation Details. Please state the exact values used for the reported results.
  4. [Reproducibility Checklist] Checklist items 4.7 and 4.10 are marked “yes,” but the paper does not describe any seed-setting procedure or state the number of algorithm runs used to compute each reported number. Please either correct the checklist or add the missing information.
  5. [Ablative Study, Table 2] The component ablation uses a random-selection baseline, which gives large apparent gains (+13.3 for CMR, +5.7 for FPV) but is a very weak baseline. A non-random single-stage reasoner or a non-hierarchical LLM selector would better support the claim that the hierarchical decomposition itself, rather than merely using an LLM, is what helps.

Circularity Check

0 steps flagged

No significant circularity: claims are evaluated against external benchmarks and no fitted equation is relabeled as a prediction.

full rationale

The paper's derivation chain is empirical rather than definitional. M1 parses the query, M2 generates candidate trajectories with off-the-shelf GroundingDINO/SAM2, and M3 selects a trajectory via an LLM prompted with serialized trajectories, a standard affine camera-motion estimate (Eq. 1), and a CLIP cosine-similarity pose check (Eq. 2). None of these components is fitted to the reported J&F numbers, and the SOTA claims are measured with official benchmark evaluation code against external datasets. The hyperparameters tau=15 and window=3 are chosen via ablations on MeViS validation (Tables 4-5), which is a benchmark-tuning caveat rather than a circular reduction of the reported results to the inputs. The self-citations in the Related Work and Introduction (e.g., Liu et al. 2024; Ying et al. 2023; Zhong et al. 2025) are background and not load-bearing for the central claim. The appendix prompt in Fig. 6(b) instructing the LLM to output only the single best candidate ID is inconsistent with the described conditional FPV stage and with K>1 queries; however, this is a reproducibility/specification contradiction, not a circularity in the derivation. Therefore no circular step is present.

Axiom & Free-Parameter Ledger

9 free parameters · 7 axioms · 0 invented entities

The framework contains no new physical or mathematical entities. Its central claim rests on pretrained perception models, a zero-shot LLM reasoner, and several hand-set/tuned thresholds; the depth-priority heuristic is the least-supported invention.

free parameters (9)
  • Keyframe sampling interval tau = 15
    Tuned on the MeViS validation set (Table 4); peak J&F 52.4 at tau=15.
  • Verification window size = 3
    Tuned on the MeViS validation set (Table 5); window 3 chosen as efficiency/accuracy trade-off.
  • Association IoU threshold theta_iou = 0.6
    Hand-set threshold for predictive association criterion; not ablated.
  • Association centroid distance threshold theta_dist = 50 pixels
    Hand-set threshold for predictive association criterion; not ablated.
  • Discriminative keyframe count k for pose verification = 3
    Hand-selected for CLIP pose verification.
  • GroundingDINO box threshold = 0.3
    Hand-set object detection confidence threshold.
  • GroundingDINO NMS IoU threshold = 0.4
    Hand-set non-maximum suppression threshold.
  • LLM temperature = 0.7
    Chosen to balance parsing determinism and reasoning creativity; no seed reported.
  • LLM top_p = 0.95
    Nucleus sampling threshold; chosen for output coherence.
axioms (7)
  • standard math Lucas-Kanade optical flow minimization yields a valid affine camera motion model (Eq. 1).
    Used in M3 camera motion prior; assumes small inter-frame displacement and planar scene motion.
  • domain assumption IoU and centroid distance are sufficient association metrics for trajectory formation.
    Predictive association criterion in M2.
  • domain assumption GroundingDINO and SAM2 provide reliable zero-shot detection and mask propagation.
    Whole M2 depends on these pretrained models; no failure analysis is provided.
  • domain assumption Llama-3-8B-Instruct can act as a zero-shot spatio-temporal reasoner from serialized text.
    M3's core premise; no calibration or consistency check is reported.
  • ad hoc to paper Larger pixel cardinality implies higher depth priority in overlapping regions.
    Occlusion relationship prior in M3; the cited Zhu et al. 2017 paper is about amodal segmentation, not this depth rule.
  • domain assumption CLIP cosine similarity on box crops captures pose and attribute distinctions.
    Fine-grained pose verification in M3, Eq. (2).
  • ad hoc to paper Sampling keyframes every tau=15 frames captures all target-relevant events.
    Keyframe sampling in M2; tau tuned on MeViS.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Unleashing Hierarchical Reasoning: An LLM-Driven Framework for Training-Free Referring Video Object Segmentation." pith.science (2026). https://pith.science/paper/OE7B57UL

@misc{pith2026250905751,
  author       = {Pith},
  title        = {Pith review of: Unleashing Hierarchical Reasoning: An LLM-Driven Framework for Training-Free Referring Video Object Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OE7B57UL}},
  note         = {Machine review of arXiv:2509.05751}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Referring Video Object Segmentation (RVOS) aims to segment an object of interest throughout a video based on a language description. The prominent challenge lies in aligning static text with dynamic visual content, particularly when objects exhibiting similar appearances with inconsistent motion and poses. However, current methods often rely on a holistic visual-language fusion that struggles with complex, compositional descriptions. In this paper, we propose \textbf{PARSE-VOS}, a novel, training-free framework powered by Large Language Models (LLMs), for a hierarchical, coarse-to-fine reasoning across text and video domains. Our approach begins by parsing the natural language query into structured semantic commands. Next, we introduce a spatio-temporal grounding module that generates all candidate trajectories for all potential target objects, guided by the parsed semantics. Finally, a hierarchical identification module select the correct target through a two-stage reasoning process: it first performs coarse-grained motion reasoning with an LLM to narrow down candidates; if ambiguity remains, a fine-grained pose verification stage is conditionally triggered to disambiguate. The final output is an accurate segmentation mask for the target object. \textbf{PARSE-VOS} achieved state-of-the-art performance on three major benchmarks: Ref-YouTube-VOS, Ref-DAVIS17, and MeViS.

Figures

Figures reproduced from arXiv: 2509.05751 by Bingrui Zhao, Deyin Liu, Jialie Shen, Lin Yuanbo Wu, Lu Zhang, Ruyi He, Xiangtian Fan, Ximing Li.

Figure 1
Figure 1. Figure 1: Comparison between our hierarchical reasoning [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed hierarchical, coarse-to-fine framework (PARSE-VOS). The pipeline primarily consists of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: An overview of our Spatio-Temporal Grounding Module. (Left) Text-Guided Instance Segmentation: We generate [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of our coarse-grained motion reason [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results on the MeViS dataset, showing our framework’s robustness and advanced contextual reasoning. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: LLM Prompts for Query Decomposition and Motion Reasoning. The figure presents the specific prompts used in our [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison between our method and the baseline DsHmp on the MeViS dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison between our method and the baseline DsHmp on the MeViS dataset. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗

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

39 extracted references · 28 canonical work pages · 3 internal anchors

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bai, Z.; He, T.; Mei, H.; Wang, P.; Gao, Z.; Chen, J.; Liu, L.; Zhang, Z.; and Shou, M. Z. 2024. One Token to Seg Them All: Language Instructed Reasoning Segmentation in Videos. In Advances in Neural Information Processing Systems

    work page 2024

  4. [4]

    Bar-Tal, O.; Ofri-Amar, D.; Fridman, R.; Kasten, Y.; and Dekel, T. 2022. Text2live: Text-driven layered image and video editing. In European Conference on Computer Vision(ECCV), volume 13675 of Lecture Notes in Computer Science, 707--723. Springer Nature Switzerland

    work page 2022

  5. [5]

    Botach, A.; Zheltonozhskii, E.; and Baskin, C. 2022. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 19777--19787

    work page 2022

  6. [6]

    Ding, H.; Liu, C.; He, S.; Jiang, X.; and Loy, C. C. 2023. MeViS : A Large-scale Benchmark for Video Segmentation with Motion Expressions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 10156--10166

    work page 2023

  7. [7]

    Gao, X.; Wu, K.; Zhang, H.; Tian, K.; Zhou, Y.; and Tu, Z. 2025 a . Automated Vehicles Should be Connected with Natural Language. arXiv preprint arXiv:2507.01059

    Pith/arXiv arXiv 2025

  8. [8]

    Gao, X.; Wu, Y.; Wang, R.; Liu, C.; Zhou, Y.; and Tu, Z. 2025 b . LangCoop: Collaborative Driving with Language. arXiv preprint arXiv:2504.13406

    Pith/arXiv arXiv 2025

  9. [9]

    Gavrilyuk, K.; Ghodrati, A.; Li, Z.; and Snoek, C. G. M. 2017. Actor and Action Video Segmentation from a Sentence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 5958--5966

    work page 2017

  10. [10]

    Gong, L.; Lin, Y.; Zhang, X.; Lu, Y.; Han, X.; Liu, Y.; Guo, S.; Lin, Y.; and Wan, H. 2024. Mobility- LLM : Learning Visiting Intentions and Travel Preferences from Human Mobility Data with Large Language Models. In Thirty-eighth Conference on Neural Information Processing Systems

    work page 2024

  11. [11]

    Gong, S.; Zhuge, Y.; Zhang, L.; Yang, Z.; Zhang, P.; and Lu, H. 2025. The Devil is in Temporal Token: High Quality Video Reasoning Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2025

  12. [12]

    He, S.; and Ding, H. 2024. Decoupling static and hierarchical motion perception for referring video segmentation. In CVPR

    work page 2024

  13. [13]

    Huang, S.; Ling, R.; Li, H.; Hui, T.; Tang, Z.; Wei, X.; Han, J.; and Liu, S. 2025. Unleashing the Temporal-Spatial Reasoning Capacity of GPT for Training-Free Audio and Language Referenced Video Object Segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence

    work page 2025

  14. [14]

    Khoreva, A.; Rohrbach, A.; and Schiele, B. 2018. Video Object Segmentation with Language Referring Expressions. In Proceedings of the Asian Conference on Computer Vision (ACCV)

    work page 2018

  15. [15]

    Kim, S.; Jin, W.; Choi, H.; Lim, S.; Kim, S.; and Yoon, H. 2025. Referring Video Object Segmentation via Language-aligned Track Selection. arXiv preprint arXiv:2412.01136

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  16. [16]

    Li, X.; Wang, J.; Xu, X.; Li, X.; Raj, B.; and Lu, Y. 2023. Robust Referring Video Object Segmentation with Cyclic Structural Consensus. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 22236--22245

    work page 2023

  17. [17]

    Liang, T.; Lin, K.-Y.; Tan, C.; Zhang, J.; Zheng, W.-S.; and Hu, J.-F. 2025. ReferDINO: Referring Video Object Segmentation with Visual Grounding Foundations. In Proceedings of the IEEE/CVF International Conference on Computer Vision

    work page 2025

  18. [18]

    Lin, L.; Wang, Y.-X.; Yu, X.; and Pang, Z. 2025. GLUS: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation . arXiv preprint arXiv:2504.07962

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    Y.; and Xie, X

    Liu, D.; Wu, L. Y.; and Xie, X. 2024. Blended Latent Diffusion under Attention Control for Real-World Video Editing. In International Conference on Pattern Recognition and Artificial Intelligence, --

    work page 2024

  20. [20]

    Liu, S.; Zeng, Z.; Ren, T.; Li, F.; Zhang, H.; Yang, J.; Li, C.; Yang, J.; Su, H.; Zhu, J.; et al. 2023. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499

    Pith/arXiv arXiv 2023

  21. [21]

    D.; and Kanade, T

    Lucas, B. D.; and Kanade, T. 1981. An Iterative Image Registration Technique with an Application to Stereo Vision . In Proceedings of the 7th International Joint Conference on Artificial Intelligence (IJCAI '81), 674--679

    work page 1981

  22. [22]

    Meta AI . 2024. The Llama 3 Herd of Models . arXiv:2404.11225

    Pith/arXiv arXiv 2024

  23. [23]

    Miao, B.; Bennamoun, M.; Gao, Y.; and Mian, A. 2023. Spectrum-guided multi-granularity referring video object segmentation. In ICCV

    work page 2023

  24. [24]

    Pan, F.; Fang, H.; Li, F.; Xu, Y.; Li, Y.; Benini, L.; and Lu, X. 2025. Semantic and Sequential Alignment for Referring Video Object Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2025

  25. [25]

    Y.; Shen, C.; and van den Hengel, A

    Qi, Y.; Wu, Q.; Anderson, P.; Wang, X.; Wang, W. Y.; Shen, C.; and van den Hengel, A. 2020. REVERIE : Remote Embodied Visual Referring Expression in Real Indoor Environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 9985--9994

    work page 2020

  26. [26]

    Qin, Z.; Han, C.; Wang, Q.; Nie, X.; Yin, Y.; and Xiankai, L. 2023. Unified 3D Segmenter as Prototypical Classifiers. In Advances in Neural Information Processing Systems (NeurIPS), volume 36, 5315--5328

    work page 2023

  27. [27]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I. 2021. Learning Transferable Visual Models From Natural Language Supervision . In Proceedings of the 38th International Conference on Machine Learning (ICML), volume 139 of Proceedings of Machine Learning Re...

    work page 2021

  28. [28]

    V.; Carion, N.; Wu, C.-Y.; Girshick, R.; Dollár, P.; and Feichtenhofer, C

    Ravi, N.; Gabeur, V.; Hu, Y.-T.; Hu, R.; Ryali, C.; Ma, T.; Khedr, H.; Rädle, R.; Rolland, C.; Gustafson, L.; Mintun, E.; Pan, J.; Alwala, K. V.; Carion, N.; Wu, C.-Y.; Girshick, R.; Dollár, P.; and Feichtenhofer, C. 2024. SAM 2: Segment Anything in Images and Videos. arXiv:2408.00714

    Pith/arXiv arXiv 2024

  29. [29]

    Seo, S.; Lee, J.-Y.; and Han, B. 2020. URVOS : Unified Referring Video Object Segmentation Network with a Large-Scale Benchmark. In Proceedings of the European Conference on Computer Vision (ECCV), 218--234

    work page 2020

  30. [30]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. In Advances in neural information processing systems, 5998--6008

    2017

  31. [31]

    Wu, J.; Jiang, Y.; Sun, P.; Yuan, Z.; and Luo, P. 2022. Language as queries for referring video object segmentation. In CVPR

    work page 2022

  32. [32]

    Yan, C.; Wang, H.; Yan, S.; Jiang, X.; Hu, Y.; Kang, G.; Xie, W.; and Gavves, E. 2024. VISA: Reasoning Video Object Segmentation via Large Language Models. arXiv preprint arXiv:2407.11325

    Pith/arXiv arXiv 2024

  33. [33]

    Yang, K.; Guo, Z.; Lin, G.; Dong, H.; Huang, Z.; Wu, Y.; Zuo, D.; Peng, J.; Zhong, Z.; Wang, X.; Guo, Q.; Jia, X.; Yan, J.; and Lin, D. 2025. TRAJECTORY-LLM : A Language-Based Data Generator for Trajectory Prediction in Autonomous Driving. In International Conference on Learning Representations

    work page 2025

  34. [34]

    Y.; Chen, H.; Liu, Y.; Fan, C.; Zhuge, Y.; and Shen, C

    Ying, K.; Zhong, Q.; Mao, W.; Wu, L. Y.; Chen, H.; Liu, Y.; Fan, C.; Zhuge, Y.; and Shen, C. 2023. CTVIS: Consistent Training for Online Video Instance Segmentation. In IEEE International Conference on Computer Vision (ICCV), --

    work page 2023

  35. [35]

    Zhang, Z.; Amiri, H.; Liu, Z.; Züfle, A.; and Zhao, L. 2023. Large Language Models for Spatial Trajectory Patterns Mining. arXiv:2310.04942

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  36. [36]

    Zheng, R.; Qi, L.; Chen, X.; Wang, Y.; Wang, K.; Qiao, Y.; and Zhao, H. 2024. ViLLa: Video Reasoning Segmentation with Large Language Model. arXiv preprint arXiv:2407.14500

    Pith/arXiv arXiv 2024

  37. [37]

    Y.; and Huang, K

    Zhong, Q.; Jiang, P.; Wang, W.; Ding, G.; Wu, L. Y.; and Huang, K. 2025. A Temporal Modeling Framework for Video Pre-Training on Video Instance Segmentation. In International Conference on Multimedia Expo (ICME), --

    work page 2025

  38. [38]

    Zhu, J.; Cheng, Z.-Q.; He, J.-Y.; Li, C.; Luo, B.; Lu, H.; Geng, Y.; and Xie, X. 2023. Tracking with human-intent reasoning. arXiv:2312.17448

    Pith/arXiv arXiv 2023

  39. [39]

    Zhu, Y.; Tian, Y.; Li, G.-T.; and Metaxas, D. N. 2017. Semantic Amodal Instance Segmentation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2128--2136

    work page 2017

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.