Pith. sign in

REVIEW 3 major objections 6 minor 68 references

TransFlow: Motion Knowledge Transfer from Video Diffusion Models to Video Salient Object Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TransFlow transfers motion knowledge from pre-trained video diffusion models to generate realistic optical flows for static images, improving video salient object detection on four benchmarks.

desk verdict A genuinely new data-synthesis recipe for video SOD with real benchmark gains, but the semantic-motion explanation is confounded by a 23x larger training set and needs a matched-scale warping control. read the letter →

arxiv 2507.19789 v1 pith:CZFPD5FJ submitted 2025-07-26 cs.CV

Suhwan Cho , Minhyeok Lee , Jungho Lee , Sunghun Yang , Sangyoun Lee This is my paper · ORCID
classification cs.CV
keywords videosalientobjectdetectionmotionknowledgetransferdiffusionmodelssynthetictrainingdataopticalflowimage-to-videogenerationscarcity
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

TransFlow sets out to solve a data bottleneck in video salient object detection: video datasets with flow annotations are scarce, while image datasets are abundant. The paper's claim is that pre-trained video diffusion models, which have learned how objects move from large-scale video, can transfer that motion knowledge by generating target frames from static images. Optical flows computed between the source image and these generated targets supply realistic motion supervision, yielding a synthetic dataset of 218,008 image-flow-mask triplets from DUTS. Trained on a mixture of this synthetic data and real video, a two-stream saliency network reaches the highest S-measure on DAVIS 2016, FBMS, DAVSOD, and ViSal, and mixed training outperforms real-only training by 3.5 points in average S-measure. The broader proposition is that generative models can act as data engines for discriminative video understanding.

What carries the argument

The machinery is a three-stage data-generation pipeline, not a new network. A source image from DUTS is the first frame; Stable Video Diffusion (a 3D-UNet latent video diffusion model) denoises 14 temporally-coherent target frames conditional on that source latent; RAFT then estimates 14 optical flow maps from source to targets. The central object is the 'semantically-aware optical flow' — flow that reflects how the salient object would move, rather than a uniform geometric warp — and it carries the argument because it turns abundant image-mask pairs into flow-supervised video training triplets. The segmentation model itself is a standard two-stream design with MiT encoders, CBAM fusion, and a progressive decoder.

What would settle it

Train the same two-stream network on 218,008 triplets built from the same DUTS images by geometric warping (spline or affine deformations) with RAFT flows, matching the mixed training schedule and all hyperparameters. If the warped-flow model matches TransFlow's S-measures (94.5, 92.6, 80.3, 96.2), the claim that diffusion-model motion priors supply the improvement would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that the motion priors inside video diffusion models are a usable source of training signal for video SOD. Instead of warping a static image to fake a second frame, TransFlow feeds the source image into Stable Video Diffusion, generates 14 target frames, and runs RAFT from source to each target. The resulting flows preserve the object's boundaries while showing coherent object-specific motion, so the ground-truth saliency mask of the source image stays perfectly aligned with both RGB and flow. With the full DUTS set this yields 218,008 triplets, about 105 times the size of DAVIS 2016, and training a two-stream RGB/flow network on a 2:1:1 mix of synthetic, DAVIS, and DAVSOD data gives S-measures of 94.5, 92.6, 80.3, and 96.2 on DAVIS 2016, FBMS, DAVSOD, and ViSal, with average S 90.9 versus 87.4 for real-only training. The paper's interpretation is that semantic motion generation, not architecture, is what unlocks the gains.

Load-bearing premise

The assumption is that the generated next frames move the salient object the way real videos do, so the calculated motion patterns carry true semantic cues instead of just adding far more training examples.

Editorial extensions

If this is right

  • If the transfer claim is right, flow-guided video SOD can be trained almost entirely from static image data, with real video used as a smaller complement.
  • The same recipe could supply motion supervision to other flow-guided tasks that suffer from scarce video annotations, such as video object segmentation and video instance segmentation.
  • Increasing synthetic training volume should continue to help, since the paper shows Jaccard scores rise with the share of synthetic data even though convergence slows.
  • The bottleneck in current video SOD is data abundance rather than model sophistication, so future effort should go into data synthesis.
  • Combining synthetic flows with real videos is better than either alone, so synthetic data is a complement, not a replacement.

Reading between the lines

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

  • If the motion-prior transfer generalizes, the same pipeline could supply motion supervision to other flow-guided tasks, and models could be adapted to new domains by generating flows from any image collection.
  • A natural next experiment would compare against geometric warping at equal training-set size, which would separate the contribution of semantic motion from the contribution of the roughly 23x larger training set.
  • The paper's limitations section notes checkerboard flow artifacts from patch-level copying and the inability to generate transformed masks because only RGB frames are synthesized; jointly generating images and masks would extend the approach.
  • One could test whether the diffusion prior is essential by replacing it with a flow-prediction network trained on real videos to hallucinate flows for static images, isolating what the generative prior specifically adds.
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

3 major / 6 minor

Summary. The paper introduces TransFlow, a training-data generation pipeline for video salient object detection (SOD). For each static image from the DUTS dataset, Stable Video Diffusion generates 14 target frames; RAFT optical flow is then computed between the source image and each generated target, and the triplet (source RGB, flow, source saliency mask) is used to train a two-stream SOD network. The authors create DUTS-Video with 218,008 triplets, combine them with real DAVIS-2016 and DAVSOD training data at a 2:1:1 ratio, and report improved S-measure, F-measure, and M-measure on four benchmarks (DAVIS-2016, FBMS, DAVSOD, ViSal). The central claim is that semantic motion priors from video diffusion models, rather than geometric warping or data volume alone, drive these improvements.

Significance. If the causal claim were established, TransFlow would be a practically useful cross-paradigm knowledge-transfer result: it would show that motion priors learned by generative video models can be harvested to train discriminative video-understanding models, while also providing a large, publicly reproducible synthetic training set. The paper's strengths include the clear and simple pipeline, the use of standard external benchmarks and metrics, an ablation that separates real, synthetic, and mixed training, and a public code link. I also note that the evaluation is not circular: no fitted constant or self-cited equation re-enters the benchmark results, and the hand-selected hyperparameters were not tuned on the test sets. However, the significance is currently conditional because the experiments do not separate the effect of semantic motion from the effect of a roughly 23-fold increase in training triplets and 255-fold increase in visual contexts. The paper's own statement in Section 4.3 that 'data abundance, rather than model sophistication, represents the primary bottleneck' further supports a scale-driven reading of the results.

major comments (3)
  1. [Section 3.2, Table 3, Figure 8] The central causal claim—that semantic motion from video diffusion, rather than data volume, drives the improvement—is not isolated. The 'Synthetic' row of Table 3 contains 218,008 triplets generated from 15,572 DUTS images, whereas the 'Real' row contains only about 9,262 triplets (2,079 from DAVIS-2016 plus 7,183 from DAVSOD). No baseline that generates target frames from the same DUTS images via affine or spline warping, at matched triplet count and mixture ratio, is reported. The authors' own statement in Section 4.3 that 'data abundance, rather than model sophistication, represents the primary bottleneck' and the monotonic improvement in Figure 8 with increasing synthetic-data ratio are both entirely consistent with a scale-driven explanation. The missing matched-scale geometric-warping control is load-bearing because the paper's novelty is specifically semantic motion generation versus geometric motion generation. Please add such a control, or explicitly reframe the contribution as large-scale synthetic data generation without the semantic-motion attribution.
  2. [Table 2] The state-of-the-art comparison in Table 2 is uncontrolled for training data volume. TransFlow is trained with 218,008 additional synthetic triplets, whereas the listed baselines are evaluated under their original training protocols. The benchmark numbers therefore do not identify the contribution of motion knowledge transfer; they show that a system trained with a large synthetic dataset outperforms systems trained with less or different data. To support the claimed contribution, the authors should either retrain the strongest baselines with the same DUTS-Video data or with the same total number of training triplets, or clearly state that Table 2 is a system-level comparison and move the mechanistic evidence entirely to a controlled ablation.
  3. [Section 4.5, Figures 5–6] The evidence that the generated flows are 'semantically aware' is qualitative. Figures 5 and 6 show selected examples, but selected examples cannot rule out that the synthetic flows are merely visually plausible or diverse without carrying the claimed object-level semantic structure. Please provide a quantitative characterization of the synthetic flows—for example, alignment of motion boundaries with the salient-object mask, comparison of flow magnitude statistics with real video flows, or a transfer experiment to an independent motion-guided task. Without such a measure, the mechanism claim remains untested, even if the downstream benchmark numbers are correct.
minor comments (6)
  1. [Section 3.2] The phrase 'N Tcomplete training triplets' is a typo; it should read 'N × T complete training triplets'.
  2. [Throughout] 'DA VIS', 'DA VSOD', and 'V AE' appear with inconsistent spacing; please use 'DAVIS', 'DAVSOD', and 'VAE' consistently.
  3. [Equation (1)] The notation z(T)s is used without a precise definition of the temporal replication operation; please define it explicitly in the text.
  4. [Table 2 and Section 4.1] The averaging protocol for the reported metrics is not specified. Please state whether S-measure, F-measure, and M-measure are averaged per frame or per video, particularly for ViSal, which has only 193 annotated frames.
  5. [Figure 2] The left branch labeled 'Conventional Training Set' and the right branch labeled 'Proposed Training Set' would benefit from a clearer explanation of how the 'Pairing' and 'Stack' steps map to the triplets used in Table 3.
  6. [Section 4.5, Figure 10] The limitation about checkerboard artifacts and the lack of mask generation is acknowledged, but the paper does not indicate how frequently such artifacts occur or whether they affect training; a short quantitative estimate would make the limitation more actionable.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the method is an empirical data-generation pipeline evaluated on external benchmarks, with no fitted parameter or self-cited theorem re-entering the reported results.

full rationale

The derivation chain is self-contained as an empirical system: static DUTS images are fed to Stable Video Diffusion (an external pretrained model) to synthesize target frames, RAFT (also external) computes optical flows, and a standard two-stream network is trained on the resulting image-flow-mask triplets and evaluated on DAVIS 2016, FBMS, DAVSOD, and ViSal. No equation in the paper defines the reported metric in terms of the method's own fitted values; the hyperparameters (guidance scale, mixture ratio, resolution) were not fitted to the test benchmarks. The self-citations in Sections 2 and 4 appear as related work or baseline methods and are not load-bearing for the central claim, which depends on the externally published SVD motion prior rather than on an unverified uniqueness theorem or ansatz imported from the authors' prior work. The closest concern is experimental rather than circular: Table 3 lacks a matched-scale geometric-warping control, so the superiority of diffusion-generated motion over geometric motion is not fully isolated from data volume and visual diversity, a point the paper itself acknowledges by attributing gains to data abundance in Section 4.3. That confound affects causal attribution, not circularity, since all benchmark numbers are externally measured. The limitation discussion in Section 4.5 also states real limitations without deriving the core result from them. Overall, the paper's reported results are externally falsifiable and not equivalent to its inputs by construction.

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

The method rests on three domain assumptions about the generative model, the flow estimator, and the use of source masks; none are formalized or verified independently in the paper. No free parameters are fitted to benchmark targets; hyperparameters such as SVD sampling settings and the 2:1:1 mixture ratio are disclosed in the text. No new physical entities are introduced.

assumptions (3)
  • domain assumption Stable Video Diffusion encodes semantic motion priors from large-scale video data.
    Invoked in Section 1 and Section 3.1; the paper does not validate this independently beyond downstream SOD performance.
  • domain assumption RAFT optical flow between a source image and a SVD-generated target frame approximates real video flow sufficiently for SOD training.
    Section 3.2 uses RAFT on source-target pairs without any quantitative check against real flows.
  • domain assumption Saliency masks from DUTS remain valid supervision when paired with source-to-target flows.
    Section 3.2 constructs triplets using the source mask and flow from source to generated target; this assumes the generated target does not alter source semantics relevant to saliency.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TransFlow: Motion Knowledge Transfer from Video Diffusion Models to Video Salient Object Detection." pith.science (2026). https://pith.science/paper/CZFPD5FJ

@misc{pith2026250719789,
  author       = {Pith},
  title        = {Pith review of: TransFlow: Motion Knowledge Transfer from Video Diffusion Models to Video Salient Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZFPD5FJ}},
  note         = {Machine review of arXiv:2507.19789}
}
read the original abstract

Video salient object detection (SOD) relies on motion cues to distinguish salient objects from backgrounds, but training such models is limited by scarce video datasets compared to abundant image datasets. Existing approaches that use spatial transformations to create video sequences from static images fail for motion-guided tasks, as these transformations produce unrealistic optical flows that lack semantic understanding of motion. We present TransFlow, which transfers motion knowledge from pre-trained video diffusion models to generate realistic training data for video SOD. Video diffusion models have learned rich semantic motion priors from large-scale video data, understanding how different objects naturally move in real scenes. TransFlow leverages this knowledge to generate semantically-aware optical flows from static images, where objects exhibit natural motion patterns while preserving spatial boundaries and temporal coherence. Our method achieves improved performance across multiple benchmarks, demonstrating effective motion knowledge transfer.

Figures

Figures reproduced from arXiv: 2507.19789 by the authors.

Figure 1
Figure 1. Motion generation comparison: (a) geometric transfor [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Training data generation pipeline. Static images are transformed into video sequences using diffusion models, then optical flows [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Two-stream network architecture for video SOD. The [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison under challenging scenarios. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Aligned qualitative comparison between real video data and our synthetic data. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: t-SNE comparison of feature distributions from real and [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: Qualitative results on the general video cases. [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Visualization of the limitations of our simulated flows. [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 58 canonical work pages

  1. [1]

    Frequency-tuned salient region de- tection

    Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region de- tection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597–1604. IEEE, 2009. v

    work page 2009

  2. [2]

    Stem-seg: Spatio-temporal em- beddings for instance segmentation in videos

    Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal- Taix´e, and Bastian Leibe. Stem-seg: Spatio-temporal em- beddings for instance segmentation in videos. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16 , pages 158–177. Springer, 2020. ii

    work page 2020

  3. [3]

    Stable video diffusion: Scaling latent video diffusion models to large datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. ii, iii

    arXiv 2023

  4. [4]

    A thin-plate spline and the decomposition of deformations

    Fred L Bookstein and WDK Green. A thin-plate spline and the decomposition of deformations. Mathematical Methods in Medical Imaging, 2(14-28):3, 1993. ii

    work page 1993

  5. [5]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luh- man, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai. com/research/video- generation-models-as-world-simulators, 3, 2024. ii

    work page 2024

  6. [6]

    Video salient object detection via contrastive features and attention modules

    Yi-Wen Chen, Xiaojie Jin, Xiaohui Shen, and Ming-Hsuan Yang. Video salient object detection via contrastive features and attention modules. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1320–1329, 2022. v

    work page 2022

  7. [7]

    Xmem: Long- term video object segmentation with an atkinson-shiffrin memory model

    Ho Kei Cheng and Alexander G Schwing. Xmem: Long- term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vi- sion, pages 640–658. Springer, 2022. ii

    work page 2022

  8. [8]

    Rethink- ing space-time networks with improved memory coverage for efficient video object segmentation

    Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethink- ing space-time networks with improved memory coverage for efficient video object segmentation. Advances in Neural Information Processing Systems, 34:11781–11794, 2021. ii

    work page 2021

Show all 68 references
  1. [9]

    Global contrast based salient region detection

    Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and ma- chine intelligence, 37(3):569–582, 2014. ii

    2014

  2. [10]

    Tack- ling background distraction in video object segmentation

    Suhwan Cho, Heansung Lee, Minhyeok Lee, Chaewon Park, Sungjun Jang, Minjung Kim, and Sangyoun Lee. Tack- ling background distraction in video object segmentation. In European Conference on Computer Vision, pages 446–462. Springer, 2022. ii

    2022

  3. [11]

    Treating mo- tion as option to reduce motion dependency in unsupervised video object segmentation

    Suhwan Cho, Minhyeok Lee, Seunghoon Lee, Chaewon Park, Donghyeong Kim, and Sangyoun Lee. Treating mo- tion as option to reduce motion dependency in unsupervised video object segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, page...

    2023

  4. [12]

    Dual pro- totype attention for unsupervised video object segmentation

    Suhwan Cho, Minhyeok Lee, Seunghoon Lee, Dogyoon Lee, Heeseung Choi, Ig-Jae Kim, and Sangyoun Lee. Dual pro- totype attention for unsupervised video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19238–19247, 2024. ii

    2024

  5. [13]

    3d u-net: learn- ing dense volumetric segmentation from sparse annota- tion

    ¨Ozg¨un C ¸ ic ¸ek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learn- ing dense volumetric segmentation from sparse annota- tion. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, At...

    2016

  6. [14]

    Mevis: A large-scale benchmark for video segmentation with motion expressions

    Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Proceed- ings of the IEEE/CVF international conference on computer vision, pages 2694–2703, 2023. i

    2023

  7. [15]

    Mose: A new dataset for video object segmentation in complex scenes

    Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. In Proceed- ings of the IEEE/CVF international conference on computer vision, pages 20224–20234, 2023. i, viii

    2023

  8. [16]

    The pascal visual object classes (voc) challenge

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303–338, 2010. ii

    2010

  9. [17]

    Structure-measure: A new way to evaluate foreground maps

    Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548–4557, 2017. v

    2017

  10. [18]

    Shifting more attention to video salient object detection

    Deng-Ping Fan, Wenguan Wang, Ming-Ming Cheng, and Jianbing Shen. Shifting more attention to video salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8554–8564,

  11. [19]

    Bidirectionally learning dense spatio-temporal feature prop- agation network for unsupervised video object segmentation

    Jiaqing Fan, Tiankang Su, Kaihua Zhang, and Qingshan Liu. Bidirectionally learning dense spatio-temporal feature prop- agation network for unsupervised video object segmentation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3646–3655, 2022. v

    2022

  12. [20]

    Temporally efficient gabor transformer for unsu- pervised video object segmentation

    Jiaqing Fan, Tiankang Su, Kaihua Zhang, Bo Liu, and Qing- shan Liu. Temporally efficient gabor transformer for unsu- pervised video object segmentation. In Proceedings of the 31st ACM International Conference on Multimedia , pages 3394–3402, 2023. v

    2023

  13. [21]

    Pyramid constrained self- attention network for fast video salient object detection

    Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming- Ming Cheng, and Shao-Ping Lu. Pyramid constrained self- attention network for fast video salient object detection. In Proceedings of the AAAI conference on artificial intelli- gence, pages 10869–10876, 2020. v

    2020

  14. [22]

    Sparsectrl: Adding sparse controls to text-to-video diffusion models

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330–348. Springer, 2025. ii

    2025

  15. [23]

    Semantic contours from inverse detectors

    Bharath Hariharan, Pablo Arbel ´aez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on com- puter vision, pages 991–998. IEEE, 2011. ii

    2011

  16. [24]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. vii

    2016

  17. [25]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. iii

    2022 arXiv

  18. [26]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. ii

    2020

  19. [27]

    Lvos: A benchmark for long-term video object segmentation

    Lingyi Hong, Wenchao Chen, Zhongying Liu, Wei Zhang, Pinxue Guo, Zhaoyu Chen, and Wenqiang Zhang. Lvos: A benchmark for long-term video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13480–13492, 2023. viii

    2023

  20. [28]

    Simulflow: Simultaneously extracting feature and identifying target for unsupervised video object segmen- tation

    Lingyi Hong, Wei Zhang, Shuyong Gao, Hong Lu, and Wen- Qiang Zhang. Simulflow: Simultaneously extracting feature and identifying target for unsupervised video object segmen- tation. In Proceedings of the 31st ACM International Con- ference on Multimedia, pages 7481–7490, 2023. v

    2023

  21. [29]

    Full-duplex strategy for video object segmentation

    Ge-Peng Ji, Keren Fu, Zhe Wu, Deng-Ping Fan, Jianbing Shen, and Ling Shao. Full-duplex strategy for video object segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4922–4933, 2021. ii, v

    2021

  22. [30]

    Casnet: A cross-attention siamese net- work for video salient object detection

    Yuzhu Ji, Haijun Zhang, Zequn Jie, Lin Ma, and QM Jonathan Wu. Casnet: A cross-attention siamese net- work for video salient object detection. IEEE transactions on neural networks and learning systems, 32(6):2676–2690,

  23. [31]

    Auto-encoding variational bayes

    Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. iii

    2013 arXiv

  24. [32]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  25. [33]

    Pika 1.0, 2023

    Pika Labs. Pika 1.0, 2023. https://pika.art/. ii

    2023

  26. [34]

    Unsupervised video object seg- mentation via prototype memory network

    Minhyeok Lee, Suhwan Cho, Seunghoon Lee, Chaewon Park, and Sangyoun Lee. Unsupervised video object seg- mentation via prototype memory network. In Proceedings of the IEEE/CVF winter conference on applications of com- puter vision, pages 5924–5934, 2023. ii

    2023

  27. [35]

    Guided slot attention for unsupervised video object segmentation

    Minhyeok Lee, Suhwan Cho, Dogyoon Lee, Chaewon Park, Jungho Lee, and Sangyoun Lee. Guided slot attention for unsupervised video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3807–3816, 2024. ii

    2024

  28. [36]

    Mo- tion guided attention for video salient object detection

    Haofeng Li, Guanqi Chen, Guanbin Li, and Yizhou Yu. Mo- tion guided attention for video salient object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7274–7283, 2019. ii

    2019

  29. [37]

    Transformer-based visual segmenta- tion: A survey

    Xiangtai Li, Henghui Ding, Haobo Yuan, Wenwei Zhang, Jiangmiao Pang, Guangliang Cheng, Kai Chen, Ziwei Liu, and Chen Change Loy. Transformer-based visual segmenta- tion: A survey. IEEE transactions on pattern analysis and machine intelligence, 2024. i

    2024

  30. [38]

    Movideo: Motion-aware video generation with diffusion model

    Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. Movideo: Motion-aware video generation with diffusion model. In European Con- ference on Computer Vision , pages 56–74. Springer, 2025. ii

    2025

  31. [39]

    Making a case for 3d convolutions for object segmentation in videos

    Sabarinath Mahadevan, Ali Athar, Aljo ˇsa O ˇsep, Sebastian Hennen, Laura Leal-Taix´e, and Bastian Leibe. Making a case for 3d convolutions for object segmentation in videos. arXiv preprint arXiv:2008.11516, 2020. ii, v

    2008 arXiv

  32. [40]

    Segmentation of moving objects by long term video analysis

    Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. IEEE trans- actions on pattern analysis and machine intelligence, 36(6): 1187–1200, 2013. iv, v

    2013

  33. [41]

    Fast video object segmentation by reference- guided mask propagation

    Seoung Wug Oh, Joon-Young Lee, Kalyan Sunkavalli, and Seon Joo Kim. Fast video object segmentation by reference- guided mask propagation. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 7376–7385, 2018. ii

    2018

  34. [42]

    Video object segmentation using space-time memory networks

    Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9226–9235, 2019. ii

    2019

  35. [43]

    Multi-scale interactive network for salient object detection

    Youwei Pang, Xiaoqi Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 9413–9422, 2020. v

    2020

  36. [44]

    Hierarchical feature align- ment network for unsupervised video object segmentation

    Gensheng Pei, Fumin Shen, Yazhou Yao, Guo-Sen Xie, Zhenmin Tang, and Jinhui Tang. Hierarchical feature align- ment network for unsupervised video object segmentation. In European Conference on Computer Vision , pages 596–

  37. [45]

    A benchmark dataset and evaluation methodology for video object segmentation

    Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, page...

  38. [46]

    Learning video object segmentation from static images

    Federico Perazzi, Anna Khoreva, Rodrigo Benenson, Bernt Schiele, and Alexander Sorkine-Hornung. Learning video object segmentation from static images. In Proceedings of the IEEE conference on computer vision and pattern recog- nition, pages 2663–2672, 2017. ii

    2017

  39. [47]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. ii

    2022

  40. [48]

    Gen-2, 2023

    runway. Gen-2, 2023. https://runwayml.com/ research/gen-2. ii

    2023

  41. [49]

    Kernelized memory network for video object segmentation

    Hongje Seong, Junhyuk Hyun, and Euntai Kim. Kernelized memory network for video object segmentation. In Com- puter Vision–ECCV 2020: 16th European Conference, Glas- gow, UK, August 23–28, 2020, Proceedings, Part XXII 16 , pages 629–645. Springer, 2020. ii

    2020

  42. [50]

    Hierarchical mem- ory matching network for video object segmentation

    Hongje Seong, Seoung Wug Oh, Joon-Young Lee, Seong- won Lee, Suhyeon Lee, and Euntai Kim. Hierarchical mem- ory matching network for video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12889–12898, 2021. ii

    2021

  43. [51]

    Hierarchical image saliency detection on extended cssd

    Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transac- tions on pattern analysis and machine intelligence , 38(4): 717–729, 2015. ii

    2015

  44. [52]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. ii

    2010 arXiv

  45. [53]

    Improved techniques for training score-based generative models

    Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438–12448, 2020. iii

    2020

  46. [54]

    Unsupervised video object segmentation with online adversarial self-tuning

    Tiankang Su, Huihui Song, Dong Liu, Bo Liu, and Qingshan Liu. Unsupervised video object segmentation with online adversarial self-tuning. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 688–698,

  47. [55]

    A unified transformer frame- work for group-based segmentation: Co-segmentation, co- saliency detection and video salient object detection

    Yukun Su, Jingliang Deng, Ruizhou Sun, Guosheng Lin, Hanjing Su, and Qingyao Wu. A unified transformer frame- work for group-based segmentation: Co-segmentation, co- saliency detection and video salient object detection. IEEE Transactions on Multimedia, 26:313–325, 2023. v

    2023

  48. [56]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part II 16, pages 402–419. Springer,

    2020

  49. [57]

    Learning to de- tect salient objects with image-level supervision

    Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to de- tect salient objects with image-level supervision. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 136–145, 2017. iv, v

    2017

  50. [58]

    Consistent video saliency using local gradient flow optimization and global refinement

    Wenguan Wang, Jianbing Shen, and Ling Shao. Consistent video saliency using local gradient flow optimization and global refinement. IEEE Transactions on Image Processing, 24(11):4185–4196, 2015. v

    2015

  51. [59]

    Video salient object detection via fully convolutional networks

    Wenguan Wang, Jianbing Shen, and Ling Shao. Video salient object detection via fully convolutional networks. IEEE Transactions on Image Processing, 27(1):38–49, 2017. ii

    2017

  52. [60]

    F3net: fusion, feedback and focus for salient object detection

    Jun Wei, Shuhui Wang, and Qingming Huang. F3net: fusion, feedback and focus for salient object detection. In Proceed- ings of the AAAI conference on artificial intelligence , pages 12321–12328, 2020. v

    2020

  53. [61]

    Cbam: Convolutional block attention module

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. iv

    2018

  54. [62]

    Segformer: Simple and efficient design for semantic segmentation with transform- ers

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transform- ers. Advances in neural information processing systems, 34: 12077–12090, 2021. iv, viii

    2021

  55. [63]

    Learning motion-appearance co- attention for zero-shot video object segmentation

    Shu Yang, Lu Zhang, Jinqing Qi, Huchuan Lu, Shuo Wang, and Xiaoxing Zhang. Learning motion-appearance co- attention for zero-shot video object segmentation. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 1564–1573, 2021. ii

    2021

  56. [64]

    Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory

    Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. ii

    2023 arXiv

  57. [65]

    I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models

    Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. ii

    2023 arXiv

  58. [66]

    Suppress and balance: A simple gated net- work for salient object detection

    Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated net- work for salient object detection. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part II 16 , pages 35–51. Springer,

    2020

  59. [67]

    Learning discriminative feature with crf for unsupervised video object segmentation

    Mingmin Zhen, Shiwei Li, Lei Zhou, Jiaxiang Shang, Haoan Feng, Tian Fang, and Long Quan. Learning discriminative feature with crf for unsupervised video object segmentation. In Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings,...

    2020

  60. [68]

    Motion-attentive transition for zero-shot video object segmentation

    Tianfei Zhou, Shunzhou Wang, Yi Zhou, Yazhou Yao, Jianwu Li, and Ling Shao. Motion-attentive transition for zero-shot video object segmentation. In Proceedings of the AAAI conference on artificial intelligence , pages 13066– 13073, 2020. ii

    2020

Pith tools

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