Pith. sign in

REVIEW 3 major objections 6 minor 58 references

DepthFlow: Exploiting Depth-Flow Structural Correlations for Unsupervised Video Object Segmentation

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 The paper claims that synthetic optical flows derived from single-image depth maps preserve the structural cues video object segmentation needs, and that training a simple two-stream encoder-decoder on them beats existing methods on every…

desk verdict A genuinely useful data-generation trick for unsupervised VOS, with a missing control that leaves the claimed mechanism under-supported. read the letter →

arxiv 2507.19790 v1 pith:4PCEWF4D submitted 2025-07-26 cs.CV

Suhwan Cho , Minhyeok Lee , Jungho Lee , Donghyeong Kim , Sangyoun Lee This is my paper · ORCID
classification cs.CV
keywords unsupervisedvideoobjectsegmentationsyntheticdatagenerationdepth-to-flowconversionopticalflowscarcitytwo-streamnetworksalientdetectionencoder-decoderarchitecture
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

DepthFlow addresses the training-data bottleneck in unsupervised video object segmentation by synthesizing optical flow maps from single images. The paper's core claim is that VOS models need structural information in flow maps, such as object boundaries and foreground-background separation, rather than geometrically accurate motion, and that this structure is mirrored in depth. The method estimates a depth map per image, normalizes it, and applies random reversal, shifting, and scaling to produce a two-channel motion map rendered in the standard flow color code. This converts 10,553 image-mask pairs from DUTS into over 15,000 image-flow-mask triplets (DUTSv2), and training a simple two-stream encoder-decoder on this data together with DAVIS 2016 yields state-of-the-art results on DAVIS 2016, FBMS, YouTube-Objects, Long-Videos, DAVSOD, and ViSal. The paper's takeaway is that data diversity and structural fidelity matter more than geometric accuracy of flow or architectural sophistication.

What carries the argument

The key machinery is the depth-to-flow conversion pipeline defined by equations (3) through (8): a min-max normalized depth map $D$, random depth reversal $M_1 = 2r(1-D) + 2(1-r)D - 1$ with $r$ drawn from $\{0,1\}$, random additive shift $M_2 = M_1 + s$ with $s$ drawn from $[-1,1]$, random multiplicative scaling $M_3 = \alpha M_2$ with $\alpha$ drawn from $[0,1]$, per-component normalization by the maximum absolute value, and finally the standard UV-to-RGB flow visualization mapping $F = \mathrm{UV2RGB}(M)$. This mechanism transforms any single image and its segmentation mask into a complete image-flow-mask training triplet, producing the DUTSv2 dataset from DUTS. The load-bearing element is that these transformations preserve depth structure, meaning relative ordering, boundaries, and foreground scale, without needing to model real motion, which is why the synthetic flow can stand in for RAFT-computed flow during training.

What would settle it

Train the exact DepthFlow protocol but replace the synthetic flow map with the normalized depth map (or its reversed, shifted, and scaled versions) as the second encoder input, keeping all other conditions fixed; if the depth-input model matches or exceeds the flow-input model on DAVIS 2016, then the flow visualization step is not contributing beyond depth structure, contradicting the method's stated mechanism. Alternatively, evaluate on a curated set of videos where the salient object is coplanar with the background so that the depth map carries no boundary signal; the model should fail to segment those objects if depth structure is the sole source of the motion cue.

Watch

Extended reading notes

Core claim

The central discovery is the depth-to-flow structural correlation: a normalized depth map, when randomly flipped in sign, shifted, and scaled, yields a plausible optical flow field that preserves the spatial cues needed to separate the salient object from its background. The paper argues that two-stream VOS networks operate per-frame and depend on these structural motion cues rather than on precise motion vectors. Therefore the synthetic flows, despite being physically implausible as motion, serve as effective training signals; feeding them to a standard RGB-plus-flow encoder-decoder with CBAM modules produces a model that, trained on the synthetic DUTSv2 dataset mixed with DAVIS 2016, outperforms prior two-stream and hybrid approaches on all public benchmarks used. The ablation shows that training on synthetic data alone beats training on the limited real video data on FBMS, YouTube-Objects, and Long-Videos, and that the synthetic data adds the most gains where domain shift is largest.

Load-bearing premise

The claim rests on the assumption that the structural information in a depth map survives random reversal, shifting, and scaling well enough to substitute for real optical flow during training; the paper does not test the alternative of feeding the depth map itself as the second input.

Editorial extensions

If this is right

  • The data scarcity bottleneck for two-stream unsupervised VOS is removed: any large-scale image-mask or saliency dataset can be converted into flow-augmented complete triplets, making the flow modality readily available.
  • A simple encoder-decoder network with CBAM fusion matches or exceeds more elaborate architectures on the tested benchmarks, suggesting that dataset construction can carry more weight than network design.
  • Because synthetic flows are generated from single video frames, the trained model can run without a separate optical flow estimator at inference, functioning as an image-level salient object detector while retaining the option to use real flow when available.
  • The performance gains are largest on datasets that differ most from DAVIS (FBMS, YouTube-Objects, Long-Videos), indicating synthetic diversity improves domain generalization.
  • The approach sets new state-of-the-art scores on all public benchmarks evaluated: DAVIS 2016, FBMS, YouTube-Objects, Long-Videos, DAVSOD, and ViSal.

Reading between the lines

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

  • A likely explanation for the method's success is that the random transformations turn a depth map into a family of edge-preserving scalar fields; the network may be learning boundary and ordering cues from the gradient of the synthetic flow rather than from motion semantics. If so, the same pipeline could work with other structure-preserving single-image cues such as surface normals, segmentation
  • The paper never directly tests the alternative of feeding the normalized depth map (or its transformations) into the second encoder instead of the flow visualization. One untested consequence of the structural-correlation hypothesis is that such a depth-stream network would match DepthFlow's performance; if it does, the UV2RGB visualization step is not essential.
  • The method could be extended to train video object segmentation in a fully unsupervised way without any video data at all, since DUTS images plus their masks already yield complete triplets; this would sever the dependency on video datasets for training.
  • A failure mode implied by the depth dependency: when a video's most salient object does not differ in depth from its background (for example, a flat object on a wall, or a scene with no depth variation), the synthetic flow provides no separating structure, so the model may have to rely entirely on RGB appearance.
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. DepthFlow proposes a training-data generation method for two-stream unsupervised VOS. It estimates a monocular depth map from a single RGB image, converts it into a synthetic two-channel motion map via random reversal, shift, and scaling (Sec. 3.2), and then renders it as an RGB flow visualization. The method builds DUTSv2, a dataset of image-flow-mask triplets from DUTS image-mask pairs, and trains a simple encoder-decoder with two encoders (RGB and flow). Experiments on DAVIS 2016, FBMS, YouTube-Objects, Long-Videos, and video SOD benchmarks report state-of-the-art or competitive results. The central claim is that synthetic flows preserving depth-derived structure can substitute for real optical flow in training.

Significance. If the central claim is substantiated, the idea is valuable: it would turn large-scale image-mask datasets into training data for two-stream VOS without requiring video or real flow, alleviating a known data-scarcity bottleneck. The paper has several strengths: a clean, simple architecture; an aligned comparison with HFAN on the same backbones and hardware (Table 8); evaluation across multiple benchmarks; and release of code. The main caveat is that the unique ingredient—depth-to-flow conversion—is not isolated experimentally; without a depth-input control, the results may be explained by the depth structure alone, which would reduce the contribution to a much weaker claim about depth-based augmentation. The generalization to video SOD benchmarks is encouraging if the effect is real.

major comments (3)
  1. [Sec. 3.2, Eqs. (4)-(8); Tables 2-6] The synthetic flow is a deterministic function of the normalized depth map D: each channel is obtained by an affine re-scaling (flip, shift, scale) of D. The paper never runs the control of feeding D itself (or a depth-derived single-channel map with the same augmentations) into the second encoder of the same architecture. Since all reported gains over RGB-only training could be caused by the depth-boundary structure alone, the abstract's claim that 'synthetic flows preserve essential structural cues' and the key observation in Sec. 3.1 are not tested. Please add an ablation with depth as the second-stream input (e.g., D, or D after the same random augmentations but before the UV2RGB conversion) and compare it with the synthetic-flow input.
  2. [Table 6] The 'Real' vs 'Synthetic' comparison confounds the generation method with training-set size and domain: Real uses 2,079 frames from 30 DAVIS 2016 sequences, while Synthetic uses 15,572 images from DUTS. The conclusion that 'training exclusively on synthetic data outperforms real DAVIS training' is therefore not attributable to the synthetic flow generation. A controlled comparison is needed, e.g., generating synthetic flow from the DAVIS 2016 training frames themselves and comparing against real RAFT flow on the same frames, or subsampling DUTS to match the real-dataset size.
  3. [Abstract and Sec. 4.2] The abstract states 'we achieve new state-of-the-art performance on all public VOS benchmarks,' but the paper evaluates on DAVIS 2016, FBMS, YouTube-Objects, Long-Videos, and video SOD benchmarks (Tables 2-5) and does not report DAVIS 2017 or MOSE quantitatively (only qualitative examples in Fig. 6). Please either add the missing benchmarks or revise the claim to 'the evaluated benchmarks,' since the 'all public benchmarks' statement is a headline claim.
minor comments (6)
  1. [Sec. 3.2] The text says the generation operates independently along the x and y axes, but Eqs. (4)-(6) show a single set of r, s, and alpha; please state explicitly whether these are sampled separately for the x and y channels.
  2. [Eq. (7)] If the maximum absolute value is zero (e.g., a degenerate depth map), the normalization divides by zero; please specify the fallback behavior.
  3. [Sec. 3.4] Please state whether the real RAFT flows are also converted to RGB via Eq. (8) before being fed to the network, and if so, whether the same color mapping is used as for the synthetic flows.
  4. [Table 9] The 'without flow' comparison is not like-for-like: DepthFlow without real flow still uses synthetic flow as the second input, whereas TMO without flow uses no motion input. Please clarify this in the text or re-analyze the comparison.
  5. [Table 6] Consider adding the number of training triplets to each row of Table 6 to make the scale difference between the Real and Synthetic settings explicit.
  6. [Limitations] The Limitations paragraph does not mention that the synthetic data has no temporal consistency; even under the per-frame assumption, the absence of any motion dynamics in training data may limit the model on videos with strong motion boundaries.

Circularity Check

1 steps flagged · score 3.0 of 10

Partial self-definitional framing: synthetic flows are constructed as affine/color transforms of depth, so the 'depth-flow structure' claim is true by construction, though the benchmark results are externally validated.

  1. self definitional [Section 3.2, Eqs. (3)-(8); claimed validation in Section 4.2]
    "M1 = 2r(1 − D) + 2(1 − r)D − 1 ... M3 = α · M2 ... M = M3 / max(max(|Mx3|), max(|My3|)) ... F = UV2RGB(M) ... we first estimate a depth map from a source image and then convert it into a synthetic flow field that preserves essential structural cues."

    Equations (4)-(8) define the synthetic flow F as a deterministic per-pixel affine/color transform of the normalized depth map D, which itself is a deterministic function of the RGB image I via DPT-Hybrid. Therefore the synthetic flow's boundary and relative-order structure is depth structure by construction, not an independently measured property. The paper's 'key observation' in Sec. 3.1 that flow structure is highly correlated with depth is thus embedded in the generator, and the Sec. 4.2 statement that results 'validate that depth-to-flow conversion preserves the structural patterns necessary for effective VOS learning' restates the construction rather than testing it.

full rationale

The benchmark claims are not circular: all main results are evaluated against external public VOS and video-SOD datasets, and the training protocol is an empirical data-augmentation recipe. No fitted parameter is renamed as a prediction, and no load-bearing uniqueness theorem is imported from the authors' prior work. The self-citations (e.g., adopting a TMO-like encoder-decoder and CBAM from prior papers by the same group) are not load-bearing for the central data-generation claim. The one significant self-definitional element is that the synthetic optical flow is, by Eqs. (3)-(8), a random affine/color re-encoding of the monocular depth estimate of the same image; consequently the paper's claim that 'this structure is highly correlated with depth' is made true by construction rather than demonstrated. A depth-as-second-stream control would be needed to establish that the depth-to-flow conversion, as opposed to the depth structure itself, is the active ingredient. This is a partial circularity in the paper's interpretive framing, not in the externally validated benchmark numbers, so a moderate score of 3 is appropriate.

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

The synthetic flow generation has no fitted parameters in the usual sense, but several hand-chosen ranges and a mixing ratio are central to the data distribution. The key domain assumptions are the depth-flow structural correlation, the accuracy of DPT-Hybrid, the per-frame paradigm, and the preservation of structure through RGB flow visualization. No new entities are postulated.

free parameters (4)
  • Random shift range s = [-1, 1]
    Eq. (5): uniform shift added to all pixels to vary motion baseline; chosen by hand, not tuned to data.
  • Random scaling factor alpha = [0, 1]
    Eq. (6): controls flow magnitude; chosen by hand.
  • Real-to-synthetic mixing ratio = 1:3
    Section 3.4: fine-tuning ratio of DAVIS 2016 to DUTSv2; chosen by hand, no sensitivity analysis reported.
  • Depth-augmentation polarity probability = 0.5 (uniform binary)
    Eq. (4): random reversal r; a design choice.
assumptions (4)
  • domain assumption Depth maps from single images contain the structural cues needed to synthesize optical flow that trains VOS models.
    Core premise introduced in Sec. 3.2 and tested indirectly via Table 6; no direct comparison with using depth as the second input.
  • domain assumption DPT-Hybrid provides sufficiently accurate monocular depth estimates for the synthesis.
    Sec. 3.2 relies on a pre-trained depth model without error analysis.
  • domain assumption Per-frame processing with short-term motion cues is sufficient for unsupervised VOS.
    Sec. 3.1 assumes per-frame image-flow-mask pairs suffice; this underpins the synthetic data construction.
  • domain assumption The standard UV2RGB flow visualization preserves the structural information needed by the segmentation network.
    Eq. (8) converts two-channel motion to RGB; the paper provides no ablation of alternative encodings (e.g., two-channel flow).

how reviews work

0 comments
Cite this review

Pith. "Pith review of DepthFlow: Exploiting Depth-Flow Structural Correlations for Unsupervised Video Object Segmentation." pith.science (2026). https://pith.science/paper/4PCEWF4D

@misc{pith2026250719790,
  author       = {Pith},
  title        = {Pith review of: DepthFlow: Exploiting Depth-Flow Structural Correlations for Unsupervised Video Object Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4PCEWF4D}},
  note         = {Machine review of arXiv:2507.19790}
}
read the original abstract

Unsupervised video object segmentation (VOS) aims to detect the most prominent object in a video. Recently, two-stream approaches that leverage both RGB images and optical flow have gained significant attention, but their performance is fundamentally constrained by the scarcity of training data. To address this, we propose DepthFlow, a novel data generation method that synthesizes optical flow from single images. Our approach is driven by the key insight that VOS models depend more on structural information embedded in flow maps than on their geometric accuracy, and that this structure is highly correlated with depth. We first estimate a depth map from a source image and then convert it into a synthetic flow field that preserves essential structural cues. This process enables the transformation of large-scale image-mask pairs into image-flow-mask training pairs, dramatically expanding the data available for network training. By training a simple encoder-decoder architecture with our synthesized data, we achieve new state-of-the-art performance on all public VOS benchmarks, demonstrating a scalable and effective solution to the data scarcity problem.

Figures

Figures reproduced from arXiv: 2507.19790 by the authors.

Figure 1
Figure 1. Overview of our synthetic data generation approach. We [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of our synthetic flow generation process from estimated depth maps. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualized depth-to-flow conversion process. Aug [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Representative image-flow pairs from our DUTSv2 dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results on the DAVIS 2016 validation set. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results on the general video cases. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: t-SNE comparison of feature distributions from real and [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 41 canonical work pages

  1. [1]

    Rethinking atrous convolution for seman- tic image segmentation

    Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for seman- tic image segmentation. arXiv preprint arXiv:1706.05587 ,

    arXiv

  2. [2]

    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. vii

    work page 2022

  3. [3]

    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. i, iii

    work page 2014

  4. [4]

    Pixel-level bijective matching for video object segmentation

    Suhwan Cho, Heansung Lee, Minjung Kim, Sungjun Jang, and Sangyoun Lee. Pixel-level bijective matching for video object segmentation. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , pages 129–138, 2022. v

    work page 2022

  5. [5]

    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. v

    work page 2022

  6. [6]

    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, pages 5140–5149, 2023. ii, v, vi, vii, viii

    work page 2023

  7. [7]

    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, vi

    work page 2024

  8. [8]

    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. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 2694–2703, 2023. i

    work page 2023

Show all 58 references
  1. [9]

    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, iii, viii

    2023

  2. [10]

    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,

  3. [11]

    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. vii

    2022

  4. [12]

    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. vii

    2020

  5. [13]

    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. vi

    2016

  6. [14]

    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. ii

    2023

  7. [15]

    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, vi, vii

    2021

  8. [16]

    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,

  9. [17]

    Adam: A method for stochastic optimization

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

    arXiv

  10. [18]

    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, vi

    2023

  11. [19]

    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, vi

    2024

  12. [20]

    Iteratively selecting an easy reference frame makes unsupervised video object segmentation easier

    Youngjo Lee, Hongje Seong, and Euntai Kim. Iteratively selecting an easy reference frame makes unsupervised video object segmentation easier. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 1245–1253, 2022. ii, vi

    2022

  13. [21]

    Video object segmentation with adaptive feature bank and uncertain-region refinement

    Yongqing Liang, Xin Li, Navid Jafari, and Jim Chen. Video object segmentation with adaptive feature bank and uncertain-region refinement. Advances in Neural Informa- tion Processing Systems, 33:3430–3441, 2020. v, vi

    2020

  14. [22]

    F2net: Learning to focus on the foreground for unsupervised video object segmentation

    Daizong Liu, Dongdong Yu, Changhu Wang, and Pan Zhou. F2net: Learning to focus on the foreground for unsupervised video object segmentation. In Proceedings of the AAAI con- ference on artificial intelligence, pages 2109–2117, 2021. ii, vi

    2021

  15. [23]

    Depth-aware test-time training for zero-shot video object segmentation

    Weihuang Liu, Xi Shen, Haolun Li, Xiuli Bi, Bo Liu, Chi- Man Pun, and Xiaodong Cun. Depth-aware test-time training for zero-shot video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19218–19227, 2024. ii

    2024

  16. [24]

    See more, know more: Unsuper- vised video object segmentation with co-attention siamese networks

    Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsuper- vised video object segmentation with co-attention siamese networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3623–3632,

  17. [25]

    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, vii

    2008 arXiv

  18. [26]

    Mobilevit: light- weight, general-purpose, and mobile-friendly vision trans- former

    Sachin Mehta and Mohammad Rastegari. Mobilevit: light- weight, general-purpose, and mobile-friendly vision trans- former. arXiv preprint arXiv:2110.02178, 2021. vi

    2021 arXiv

  19. [27]

    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. v, vi

    2013

  20. [28]

    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. vi

    2019

  21. [29]

    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. vii

    2020

  22. [30]

    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–

  23. [31]

    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...

  24. [32]

    The 2017 davis challenge on video object segmentation

    Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Ar- bel´aez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. viii

    2017 arXiv

  25. [33]

    Learning object class detectors from weakly annotated video

    Alessandro Prest, Christian Leistner, Javier Civera, Cordelia Schmid, and Vittorio Ferrari. Learning object class detectors from weakly annotated video. In 2012 IEEE Conference on computer vision and pattern recognition, pages 3282–3289. IEEE, 2012. v, vi

    2012

  26. [34]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179–12188, 2021. iii

    2021

  27. [35]

    Reciprocal transformations for unsupervised video object segmentation

    Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guo- qiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455–15464, 2021. ii, vi

    2021

  28. [36]

    D2conv3d: Dynamic dilated convolutions for object segmentation in videos

    Christian Schmidt, Ali Athar, Sabarinath Mahadevan, and Bastian Leibe. D2conv3d: Dynamic dilated convolutions for object segmentation in videos. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1200–1209, 2022. ii, vi

    2022

  29. [37]

    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. i

    2015

  30. [38]

    Very deep convo- lutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. vi

    2014 arXiv

  31. [39]

    Generalizable fourier augmentation for unsupervised video object segmentation

    Huihui Song, Tiankang Su, Yuhui Zheng, Kaihua Zhang, Bo Liu, and Dong Liu. Generalizable fourier augmentation for unsupervised video object segmentation. In Proceedings of the AAAI conference on artificial intelligence , pages 4918– 4924, 2024. ii, vi

    2024

  32. [40]

    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,

  33. [41]

    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. vii

    2023

  34. [42]

    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

  35. [43]

    Video classification with channel-separated convolu- tional networks

    Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feis- zli. Video classification with channel-separated convolu- tional networks. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 5552–5561,

  36. [44]

    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. In CVPR,

  37. [45]

    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

  38. [46]

    Zero-shot video object segmenta- tion via attentive graph neural networks

    Wenguan Wang, Xiankai Lu, Jianbing Shen, David J Cran- dall, and Ling Shao. Zero-shot video object segmenta- tion via attentive graph neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9236–9245, 2019. ii, vi

    2019

  39. [47]

    Learning unsupervised video object segmentation through visual attention

    Wenguan Wang, Hongmei Song, Shuyang Zhao, Jianbing Shen, Sanyuan Zhao, Steven CH Hoi, and Haibin Ling. Learning unsupervised video object segmentation through visual attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 3064– 3...

    2019

  40. [48]

    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. vii

    2020

  41. [49]

    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. v

    2018

  42. [50]

    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. vi

    2021

  43. [51]

    Youtube-vos: A large-scale video object segmentation benchmark

    Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. i, ii, v

    2018 arXiv

  44. [52]

    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, vi

    2021

  45. [53]

    Anchor diffusion for un- supervised video object segmentation

    Zhao Yang, Qiang Wang, Luca Bertinetto, Weiming Hu, Song Bai, and Philip HS Torr. Anchor diffusion for un- supervised video object segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 931–940, 2019. ii, vii

    2019

  46. [54]

    Deep transport network for unsupervised video ob- ject segmentation

    Kaihua Zhang, Zicheng Zhao, Dong Liu, Qingshan Liu, and Bo Liu. Deep transport network for unsupervised video ob- ject segmentation. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 8781–8790,

  47. [55]

    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

  48. [56]

    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

  49. [57]

    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

  50. [613]

    ii, vi, vii

    Springer, 2022. ii, vi, vii

    2022

Pith tools

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