Pith. sign in

REVIEW 4 major objections 4 minor 46 references

FTIO: Frequent Temporally Integrated Objects

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A post-processing pipeline lifts multi-object unsupervised video segmentation to a new state of the art.

desk verdict A mostly honest post-processing tool that gets a small SOTA bump on DAVIS 2017, but its own ablation data contradict the combined-criterion story. read the letter →

arxiv 2508.16183 v1 pith:R24WSASW submitted 2025-08-22 cs.CV

Mohammad Mohammadzadeh Kalati , Farhad Maleki , Ian McQuillan This is my paper
classification cs.CV
keywords unsupervisedvideoobjectsegmentationmulti-objecttemporalconsistencyselectionopticalflowregistrationpost-processingDAVISdatasetmaskrefinement
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

The paper claims that two lightweight post-processing corrections—ranking object proposals by how often they appear (combined with mask size), then repairing temporally broken masks with optical-flow-based registration—can push unsupervised multi-object video object segmentation to a new state of the art. Applied to the DEV A baseline on the Unsupervised DAVIS 2017 dataset, FTIO reports a J&F of 75.9% on validation and 67.7% on test-dev, outperforming all previously published UVOS methods. The core insight is that zero-shot UVOS failures—objects missing from the first frame, over-segmentation, and mask erosion from deformation—can be mitigated without retraining the segmentation model.

What carries the argument

The combined object-selection score (Eq. 4) ranks objects by appearance count plus α times normalized mask area, with α=5. The three-stage temporal-consistency correction is anchored on Mask-over-Union (Eq. 5) for detecting inconsistent frames, histogram Manhattan distance (Eq. 7) for filtering occlusions, and iterative Lucas-Kanade flow-based frame registration (Eq. 8) for extracting and integrating missing mask regions.

What would settle it

Run the full FTIO pipeline on Unsupervised DAVIS 2017 validation with the Eq. 7 threshold replaced by a single fixed constant (e.g., 50) and with a perfect oracle threshold computed from ground truth. If the fixed constant yields a J&F within 0.5 points of the reported 75.9, the size-based threshold is not doing the work the paper attributes to it; if the oracle threshold yields more than 1 point higher, the heuristic is leaving measurable performance on the table.

Watch

Extended reading notes

Core claim

FTIO's central claim is that both object selection and temporal consistency can be improved without retraining the segmentation model. On the selection side, it replaces confidence-based ranking with a combined score CA = NA + 5SA, where NA is the number of frames in which the object's mask appears and SA is the cumulative normalized mask area; this recovers salient objects that the baseline misses because its image-level proposal step failed in the first frame. On the consistency side, a three-stage temporal-consistency stage detects frames where a mask shrinks abruptly using a Mask-over-Union measure against a sliding window, filters out true occlusions by comparing RGB histograms of the d

Load-bearing premise

The temporal-consistency stage assumes the histogram-based occlusion threshold (Eq. 7)—described only as 'determined based on the size of object masks' with no formula given—can reliably separate true mask loss from occlusion; if miscalibrated, the method either adds hallucinated regions or fails to repair genuine gaps.

Editorial extensions

If this is right

  • On Unsupervised DAVIS 2017, FTIO beats all previously reported multi-object UVOS methods on both validation and test-dev sets.
  • The object-selection phase alone already sets a new validation J&F (75.6%), showing that frequency-plus-size ranking is competitive with or better than size alone on this benchmark.
  • The temporal-consistency phase repairs over-segmentation and erosion, with per-object gains such as +13.3 in J for gold-fish5 and +13.2 in F for soapbox3.
  • TC propagation to adjacent frames recursively repairs missing parts across multiple consecutive frames until no inconsistencies remain.
  • The paper reports that TC can slightly hurt in cluttered scenes like 'india' due to complex scenes and misidentification, a caveat the authors document.

Reading between the lines

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

  • The same post-processing recipe may transfer to other category-agnostic proposers (any open-world segmenter) and to other multi-object UVOS datasets, especially where objects appear consistently but not necessarily in frame one.
  • The only loosely specified component is the occlusion threshold in Eq. 7; a fixed or learned threshold might match or beat the size-based heuristic, which could be tested directly.
  • Frequency-plus-size ranking could also improve supervised or semi-supervised VOS pipelines where object proposals are noisy, though the paper focuses on the unsupervised setting.
  • The runtime (6.77 s/frame on validation) makes this an offline refinement tool; a lighter registration method would be needed for online use, but the paper argues Lucas-Kanade is necessary for accuracy.
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 FTIO, a post-processing framework for unsupervised multi-object video object segmentation (UVOS). It is applied to the DEVA baseline on the DAVIS 2017 unsupervised dataset. FTIO has two phases: Object Selection (OS), which ranks object proposals by a weighted combination of appearance count and mask size (Eq. 4), and Temporal Consistency (TC), a three-stage procedure that detects temporally inconsistent frames via a Mask-over-Union statistic, filters occlusions using RGB histogram distances, and repairs missing mask parts using optical-flow-based registration. The paper reports state-of-the-art results on the DAVIS 2017 unsupervised validation (75.9 J&F) and test-dev (67.7 J&F) sets.

Significance. If the claims were fully supported, FTIO would be a useful post-processing module and a valuable study of object-selection criteria for multi-object UVOS, especially because it is evaluated on an external test-dev benchmark and the code is publicly available. The paper also includes failure-case figures and a component-wise ablation, which is helpful for reproducibility. However, the ablation data in Table 3 and the accompanying prose contradict the central claim that the combined criterion and the TC stage improve results. The test-dev set is independent and provides a measure of external validity, but the paper does not report the best configuration it found on that set. As written, the evidence does not establish the stated state-of-the-art claim or the benefit of the paper's main contributions.

major comments (4)
  1. [§3.1, Table 3 (Eq. 4)] The central claim that the combined criterion in Eq. (4) improves object selection is contradicted by the paper's own ablations. On the validation set, Single-Criterion size achieves J&F 76.33, while the combined criterion with α=5 achieves 75.94. On test-dev, the text states that appearance-count alone yields J&F 68.14, whereas the reported combined configuration yields 67.7. Thus the combined criterion is not the best on either split. Please reconcile these numbers with the abstract and Section 1 claims, or revise the central message.
  2. [§3, Table 1 vs §3.1, Table 3] There is a numerical inconsistency for the same OS configuration: Table 1 reports FTIO(OS) as 75.6 validation J&F, while Table 3 reports the Combined-Criterion α=5 OS result as 75.94 J&F. These numbers should be identical if they describe the same method. Moreover, the full FTIO pipeline (75.9 validation J&F) is lower than the size-only OS configuration (76.33), indicating that the TC stage degrades the overall validation score. This undermines the claim that adding TC improves temporal consistency. Please explain the discrepancy and provide a per-split comparison of all tested configurations.
  3. [§2.1.2, Eq. (7)] The threshold used in the occlusion-refinement stage is load-bearing: it separates true occlusion from mask inconsistency. The paper only says the threshold is "determined based on the size of object masks," with no formula, value, or implementation detail. This is not reproducible. Please specify the exact threshold function and, ideally, include a sensitivity analysis over its values.
  4. [§2.1.2, Eq. (5)] The rule for detecting inconsistent frames is ambiguous. The phrase "if more than |w|/2 times the MoU of object A at time t is higher than the standard deviation" is not a precise algorithmic statement. It is unclear whether the comparison is made inside each window, how multiple windows are aggregated, and what "more than |w|/2 times" means. This makes the first TC stage impossible to reimplement from the text. Please provide a precise formal definition, or pseudocode.
minor comments (4)
  1. [§3.1, Table 3 caption] The table header uses "Test" to mean "test category" rather than "test-dev split." Clarify in the caption that the numeric columns are the validation-set results, with test-dev results given only in prose for selected rows.
  2. [§2.1.1, Eq. (4)] The statement "Performing ablation experiments, we find α=5 appropriate" is vague. Report the criterion used to select α and note that α was tuned on the validation set; the test-dev results should be discussed explicitly in light of this selection.
  3. [Throughout] The baseline method is written inconsistently as "DEVA" in the text and "DEV A" in tables and some references. Use the conventional spelling consistently.
  4. [§3.1, Table 2] The sentence "Only the objects with at least ?? change by adding TC are shown" appears to have a rendering artifact ("??"). State the threshold explicitly.

Circularity Check

1 steps flagged · score 4.0 of 10

Validation-set headline is partly fitted because the α hyperparameter is tuned on the same validation set; the test-dev result remains an independent external benchmark.

  1. fitted input called prediction [Section 2.1.1 (Eq. 4) and Section 3.1; Table 1]
    "Hence: CA = NA + αSA (4) ... Performing ablation experiments, we find α = 5appropriate for this hyper-parameter. ... Hence, we report our final results using α = 5."

    The combined object-selection score C_A depends on the hyperparameter α, and α is selected by inspecting the DAVIS validation-set ablations reported in Section 3.1/Table 3. The same validation set is then used for the headline claim 'achieving a J&F score of 75.9% on the validation set'. Therefore the validation J&F is not an independent prediction of a fixed method; it is the metric value of a configuration chosen with access to that metric. The test-dev result (67.7%) grounds the method independently, but the validation-based state-of-the-art claim is partly fitted. The paper's own ablation also shows that α = 5 is not the best configuration on either split (size-only beats it on validation and appearance-count-only beats it on test-dev), so the tuning did not even maximize the fitted me

full rationale

The FTIO pipeline itself is defined from explicit equations and external inputs: object size (Eq. 2), appearance count (Eq. 3), the combined selection score (Eq. 4), MoU inconsistency (Eq. 5), histogram occlusion distance (Eq. 7), and optical-flow-based correction (Eq. 8). The final comparison on DAVIS test-dev is an external benchmark, and no load-bearing self-citation or imported uniqueness theorem is used. The main circularity concern is limited to the validation headline: α is tuned on the validation set and the validation score is then reported as a result, so that number is partly selected rather than predicted. The test-dev score remains independent and provides genuine support. Separately, the paper has non-circular weaknesses: its own ablation contradicts the central 'combined criterion improves selection' claim (size-only is better on validation, appearance-count-only is better on test-dev), and the occlusion threshold after Eq. 7 is described as 'determined based on the size of object masks' without a formula. These are inconsistency and reproducibility concerns, not reductions of the derivation to its inputs, so they do not raise the circularity score beyond 4.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central result rests on several hand-tuned thresholds and the assumption that frequency of appearance tracks saliency. No new physical entities are introduced.

free parameters (4)
  • alpha (α) in Eq. 4 = 5
    Weights size S_A relative to appearance count N_A; tuned on DAVIS validation via ablation experiments.
  • window length |w| = 5
    Hand-chosen; paper says at least 3 signals needed and larger windows fail with large object changes.
  • MoU majority threshold = > |w|/2 (i.e., 3 of 5)
    Hand-chosen heuristic: a frame is inconsistent if more than half the sliding windows flag it.
  • Histogram distance threshold (Eq. 7) = unspecified
    Paper says threshold is 'determined based on the size of object masks' but gives no formula.
assumptions (4)
  • domain assumption Salient objects in DAVIS appear frequently across the video sequence.
    Used to justify the appearance count term N_A in the combined criterion (Section 2.1.1).
  • ad hoc to paper A frame with a mask area that deviates from the window standard deviation in its MoU is likely inconsistent.
    The MoU detection rule is a heuristic introduced for this method; no grounding or proof of reliability is given.
  • ad hoc to paper Occlusion can be distinguished from true mask loss by comparing RGB histograms of the difference region.
    Refining stage uses Manhattan distance of histograms with an unspecified threshold to ignore occluded frames.
  • standard math Iterative Lucas-Kanade optical flow provides accurate frame-to-frame registration for mask correction.
    Lucas-Kanade is a standard technique; accuracy is assumed for the correction stage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FTIO: Frequent Temporally Integrated Objects." pith.science (2026). https://pith.science/paper/R24WSASW

@misc{pith2026250816183,
  author       = {Pith},
  title        = {Pith review of: FTIO: Frequent Temporally Integrated Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R24WSASW}},
  note         = {Machine review of arXiv:2508.16183}
}
read the original abstract

Predicting and tracking objects in real-world scenarios is a critical challenge in Video Object Segmentation (VOS) tasks. Unsupervised VOS (UVOS) has the additional challenge of finding an initial segmentation of salient objects, which affects the entire process and keeps a permanent uncertainty about the object proposals. Moreover, deformation and fast motion can lead to temporal inconsistencies. To address these problems, we propose Frequent Temporally Integrated Objects (FTIO), a post-processing framework with two key components. First, we introduce a combined criterion to improve object selection, mitigating failures common in UVOS--particularly when objects are small or structurally complex--by extracting frequently appearing salient objects. Second, we present a three-stage method to correct temporal inconsistencies by integrating missing object mask regions. Experimental results demonstrate that FTIO achieves state-of-the-art performance in multi-object UVOS. Code is available at: https://github.com/MohammadMohammadzadehKalati/FTIO

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 43 canonical work pages

  1. [1]

    Athar, S

    A. Athar, S. Mahadevan, A. O ˘sep, L. Leal-Taixé, and B. Leibe. STEm- Seg: Spatio-temporal embeddings for instance segmentation in videos. In Computer Vision – ECCV 2020, pages 158–177, 2020

    work page 2020

  2. [2]

    Brox and J

    T. Brox and J. Malik. Object segmentation by long term analysis of point trajectories. In Computer Vision – ECCV 2010 , pages 282–295, 2010

    work page 2010

  3. [3]

    Caelles, J

    S. Caelles, J. Pont-Tuset, F. Perazzi, A. Montes, K.-K. Maninis, and L. Van Gool. The 2019 DA VIS challenge on VOS: Unsupervised multi- object segmentation. arXiv:1905.00737, 2019

    arXiv 2019

  4. [4]

    H. K. Cheng and A. G. Schwing. XMem: Long-term video object seg- mentation with an Atkinson-Shiffrin memory model. In ECCV, 2022

    work page 2022

  5. [5]

    H. K. Cheng, S. W. Oh, B. Price, A. Schwing, and J.-Y . Lee. Track- ing anything with decoupled video segmentation. arXiv preprint arXiv:2309.03903, pages 1–19, Sept. 2023

    arXiv 2023

  6. [6]

    H. K. Cheng, S. W. Oh, B. Price, A. Schwing, and J.-Y . Lee. Track- ing anything with decoupled video segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages 1316–1326, Oct. 2023

    work page 2023

  7. [7]

    H. K. Cheng, S. W. Oh, B. Price, J.-Y . Lee, and A. Schwing. Putting the object back into video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3151–3161, June 2024

    work page 2024

  8. [8]

    H. K. R. Cheng. DEV A: Tracking anything with decou- pled video segmentation. https://github.com/hkchengrex/ Tracking-Anything-with-DEV A, 2023. Accessed: 2024-06-01

    work page 2023

Show all 46 references
  1. [9]

    S. Cho, M. Lee, S. Lee, D. Lee, H. Choi, I.-J. Kim, and S. Lee. Dual prototype attention for unsupervised video object segmentation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19238–19247, June 2024

    2024

  2. [10]

    Delatolas, V

    T. Delatolas, V . Kalogeiton, and D. P. Papadopoulos. Learning the what and how of annotation in video object segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6951–6961, Jan. 2024

    2024

  3. [11]

    H. Ding, C. Liu, S. He, X. Jiang, P. H. Torr, and S. Bai. MOSE: A new dataset for video object segmentation in complex scenes. In Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20224–20234, Oct. 2023

    2023

  4. [12]

    M. Gao, F. Zheng, J. J. Q. Yu, C. Shan, G. Ding, and J. Han. Deep learning for video object segmentation: a review. Artificial Intelligence Review, 56(1):457–531, 2023

    2023

  5. [13]

    S. N. Gowda, P. Eustratiadis, T. Hospedales, and L. Sevilla-Lara. Alba: Reinforcement learning for video object segmentation. In The 31st British Machine Vision Virtual Conference (BMVC), 2020

    2020

  6. [14]

    K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask R-CNN. In 2017 IEEE International Conference on Computer Vision (ICCV) , pages 2980–2988, 2017

    2017

  7. [15]

    Z. Lai, E. Lu, and W. Xie. MAST: A memory-augmented self- supervised tracker. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    2020

  8. [16]

    Le Besnerais and F

    G. Le Besnerais and F. Champagnat. Dense optical flow by iterative local window registration. In IEEE International Conference on Image Processing 2005, pages I–137, 2005

    2005

  9. [17]

    F. Li, T. Kim, A. Humayun, D. Tsai, and J. M. Rehg. Video segmen- tation by tracking many figure-ground segments. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Dec. 2013

    2013

  10. [18]

    H. Lin, R. Wu, S. Liu, J. Lu, and J. Jia. Video instance segmentation with a propose-reduce paradigm. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 1739–1748, Oct. 2021

    2021

  11. [19]

    X. Lu, W. Wang, J. Shen, Y .-W. Tai, D. J. Crandall, and S. C. H. Hoi. Learning video object segmentation from unlabeled videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    2020

  12. [20]

    Lucas and T

    B. Lucas and T. Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence (IJCAI), pages 674–679, 1981

    1981

  13. [21]

    Luiten, I

    J. Luiten, I. E. Zulfikar, and B. Leibe. UnOVOST: Unsupervised offline video object segmentation and tracking. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Mar. 2020

    2020

  14. [22]

    Meta Platforms

    I. Meta Platforms. Unsupervised video object segmentation on DA VIS 2017 (test-dev) benchmark. https://paperswithcode.com/sota/ unsupervised-video-object-segmentation-on-5, 2023. Accessed: 2024- 11-12

    2017

  15. [23]

    Meta Platforms

    I. Meta Platforms. Unsupervised video object segmentation on DA VIS 2017 (val) benchmark. https://paperswithcode.com/sota/ unsupervised-video-object-segmentation-on-4, 2023. Accessed: 2024- 10-29

    2017

  16. [24]

    P. Ochs, J. Malik, and T. Brox. Segmentation of moving objects by long term video analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(6):1187–1200, 2014

    2014

  17. [25]

    Papazoglou and V

    A. Papazoglou and V . Ferrari. Fast object segmentation in unconstrained video. In Proceedings of the IEEE International Conference on Com- puter Vision (ICCV), Dec. 2013

    2013

  18. [26]

    G. Pei, F. Shen, Y . Yao, G.-S. Xie, Z. Tang, and J. Tang. Hierarchical feature alignment network for unsupervised video object segmentation. In Computer Vision – ECCV 2022, pages 596–613, 2022

    2022

  19. [27]

    Perazzi, J

    F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016

    2016

  20. [28]

    Pont-Tuset, F

    J. Pont-Tuset, F. Perazzi, S. Caelles, P. Arbeláez, A. Sorkine-Hornung, and L. Van Gool. The 2017 DA VIS challenge on video object segmen- tation. arXiv:1704.00675, 2017

    2017 arXiv

  21. [29]

    L. Qi, J. Kuen, Y . Wang, J. Gu, H. Zhao, P. Torr, Z. Lin, and J. Jia. Open world entity segmentation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 45(7):8743–8756, 2023

    2023

  22. [30]

    D. D. A. V . Segmentation. DA VIS challenge website. https:// davischallenge.org/, 2016. Accessed: 2024-10-29

    2016

  23. [31]

    H. Song, W. Wang, S. Zhao, J. Shen, and K.-M. Lam. Pyramid dilated deeper ConvLSTM for video salient object detection. InProceedings of the European Conference on Computer Vision (ECCV), Sept. 2018

    2018

  24. [32]

    H. Song, T. Su, Y . Zheng, K. Zhang, B. Liu, and D. Liu. Generaliz- able Fourier augmentation for unsupervised video object segmentation. Proceedings of the AAAI Conference on Artificial Intelligence , 38(5): 4918–4924, Mar. 2024

    2024

  25. [33]

    T. Su, H. Song, D. Liu, B. Liu, and Q. Liu. Unsupervised video ob- ject segmentation with online adversarial self-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages 688–698, Oct. 2023

    2023

  26. [34]

    D. Tsai, M. Flagg, A. Nakazawa, and J. M. Rehg. Motion coherent tracking using multi-label MRF optimization. International Journal of Computer Vision, 100(2):190–202, 2012

    2012

  27. [35]

    Ventura, M

    C. Ventura, M. Bellver, A. Girbau, A. Salvador, F. Marques, and X. Giro-i Nieto. RVOS: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), June 2019

    2019

  28. [36]

    W. Wang, J. Shen, and F. Porikli. Saliency-aware geodesic video object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015

    2015

  29. [37]

    W. Wang, H. Song, S. Zhao, J. Shen, S. Zhao, S. C. H. Hoi, and H. Ling. Learning unsupervised video object segmentation through visual atten- tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  30. [38]

    N. Xu, L. Yang, Y . Fan, J. Yang, D. Yue, Y . Liang, B. Price, S. Cohen, and T. Huang. Youtube-VOS: Sequence-to-sequence video object seg- mentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 585–601, 2018

    2018

  31. [39]

    N. Xu, L. Yang, Y . Fan, D. Yue, Y . Liang, J. Yang, and T. Huang. YouTube-VOS: A large-scale video object segmentation benchmark, 2018

    2018

  32. [40]

    B. Yan, Y . Jiang, J. Wu, D. Wang, P. Luo, Z. Yuan, and H. Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15325–15336, June 2023

    2023

  33. [41]

    L. Yang, Y . Fan, and N. Xu. Video instance segmentation. In Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV), Oct. 2019

    2019

  34. [42]

    Y . Yuan, Y . Wang, L. Wang, X. Zhao, H. Lu, Y . Wang, W. Su, and L. Zhang. Isomer: Isomerous transformer for zero-shot video object segmentation. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision (ICCV), pages 966–976, Oct. 2023

    2023

  35. [43]

    Zhang, O

    D. Zhang, O. Javed, and M. Shah. Video object segmentation through spatially accurate and temporally dense extraction of primary object re- gions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2013

    2013

  36. [44]

    T. Zhou, J. Li, S. Wang, R. Tao, and J. Shen. MATNet: Motion- attentive transition network for zero-shot video object segmentation. IEEE Transactions on Image Processing, 29:8326–8338, 2020

    2020

  37. [45]

    T. Zhou, J. Li, X. Li, and L. Shao. Target-aware object discovery and association for unsupervised video multi-object segmentation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6985–6994, June 2021

    2021

  38. [46]

    T. Zhuo, Z. Cheng, P. Zhang, Y . Wong, and M. Kankanhalli. Unsu- pervised online video object segmentation with motion property under- standing. IEEE Transactions on Image Processing, 29:237–249, 2020

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.