AutoQ-VIS: Improving Unsupervised Video Instance Segmentation via Automatic Quality Assessment

Dim P. Papadopoulos; Kaixuan Lu; Mehmet Onurcan Kaya

REVIEW 4 major objections 5 minor 26 references

AutoQ-VIS: Improving Unsupervised Video Instance Segmentation via Automatic Quality Assessment

T0 review · 4 major / 5 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 AutoQ-VIS claims that a closed loop of pseudo-label generation and learned mask-quality scoring lets unsupervised video instance segmentation adapt from synthetic to real video, reaching 52.6 AP50 on YouTubeVIS-2019 without human annotation

desk verdict A neat quality-guided self-training idea, but the headline SOTA margin rests on a threshold chosen on the same val set; at a slightly different threshold the gain is only +0.6. read the letter →

arxiv 2508.19808 v1 pith:LEX3MMGX submitted 2025-08-27 cs.CV

Kaixuan Lu , Mehmet Onurcan Kaya , Dim P. Papadopoulos This is my paper

classification cs.CV

keywords videoinstancesegmentationunsupervisedlearningself-trainingpseudo-labelfilteringmaskqualityassessmentsynthetic-to-realdomaingapquality-guided

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

Video instance segmentation means detecting, segmenting, and tracking every object across video frames, but pixel-level masks and temporal identities are expensive to annotate by hand. The paper sets out to show that the synthetic-to-real gap—the main obstacle for unsupervised methods that train on artificially generated videos—can be closed by a self-training loop in which the model labels real videos itself and keeps only those self-generated labels (pseudo-labels) that a learned quality scorer approves. Its evidence is a rise from 48.2 to 52.6 AP50 on YouTubeVIS-2019 val over the previous unsupervised method, with no human annotations, and an ablation showing the quality scorer is worth +2.1 AP50 beyond using the model's confidence score alone. If true, unsupervised video instance segmentation can improve by recycling its own confident predictions rather than requiring optical flow or manual labels.

What carries the argument

The central object is the mask-quality predictor: a Mask Scoring R-CNN-style head that takes pixel-decoder features and a predicted mask, and regresses the mask's IoU. Combined with the detection confidence as Q_l = predicted IoU × confidence, it provides a ranking of pseudo-labels that is better correlated with true mask quality than raw confidence. This quality gate is what lets the self-training loop add real-video pseudo-labels while suppressing noisy ones. A secondary mechanism, DropLoss, removes mask-head loss terms for predictions whose maximum overlap with any ground-truth mask is below 0.01, which the ablation shows is the largest single contributor (+4.6 AP50).

What would settle it

On a sample of unlabeled YouTubeVIS-2019 train videos, generate pseudo-masks with the round-1 model and have the frozen quality predictor score them; compare those scores with the masks' true overlap against ground truth. If the rank correlation falls to the confidence score's level (about 0.42) or below, or if masks admitted at the chosen threshold have no higher average true IoU than rejected masks, then the quality gate is not selecting better masks and the core claim fails.

Watch

Extended reading notes

Core claim

AutoQ-VIS claims that unsupervised video instance segmentation improves substantially when pseudo-label generation is paired with automatic mask-quality assessment in a closed loop. A VideoMask2Former model is first trained on VideoCutLER's synthetic videos together with a mask-quality predictor inspired by Mask Scoring R-CNN. In each self-training round, the VIS model labels unlabeled real videos, the frozen quality predictor scores each pseudo-mask by its predicted IoU (average precision at 50% overlap, AP50) multiplied by the detection confidence, and only masks above a fixed threshold are fused into the training set. The predictor's score correlates more strongly with true mask overlap t

Load-bearing premise

The quality predictor, trained only on synthetic pseudo-labels, ranks real-video pseudo-masks by true overlap well enough (Spearman 0.57) that a fixed threshold admits mostly good labels and stays reliable as the VIS model's weights change across rounds; if that transfer fails, the quality filter adds noise and the reported +2.1 AP50 gain disappears.

Editorial extensions

If this is right

Unsupervised video instance segmentation can improve without optical flow or human annotations by alternating pseudo-label generation with quality-filtered dataset expansion.
A predictor trained on synthetic pseudo-labels transfers to real video well enough that its score, not the model's confidence, should be used to select pseudo-labels.
Two self-training rounds hit the optimum; a third round degrades performance by 0.6 AP50 because pseudo-label errors accumulate, so stopping early is part of the method.
The gap between the achieved 52.6 and the practical upper bound of 62.7 AP50 shows that better pseudo-label selection, not just better mask generation, is the next bottleneck.

Reading between the lines

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

The same quality-gated self-training design could transfer to other tasks that start from synthetic pretraining and unlabeled real data, such as unsupervised detection or semantic segmentation, wherever a cheap quality score can be learned.
Because the threshold is fixed and the optimal number of rounds is two, an adaptive threshold that tightens as noise accumulates might extend the loop beyond round two without the observed 0.6 AP50 drop.
The practical-limit result suggests the false positives admitted by the quality gate are less damaging than the valid pseudo-labels it rejects, so improving the quality predictor's precision at fixed recall may yield larger gains than improving the segmenter itself.
Weight resetting to the synthetic-initialized model before each round implies drift from the initial weights is harmful; one testable extension is resetting more frequently or annealing the reset strength to control drift more finely.

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.

Desk Editor's Note

read the letter

The core idea here is genuinely new and simple: train a mask IoU predictor on synthetic pseudo-labels, then use it to filter the model's own pseudo-labels on real videos during iterative self-training. That closed loop—synthetic-trained quality scoring applied to real-data self-training—is a reasonable way to attack the synthetic-to-real gap, and the paper does some things well. The ablations are mostly sensible, the oracle upper-bound experiments in Table 2 are informative, and the Spearman correlation plot (0.57 vs 0.42) gives some direct evidence that the learned quality score ranks pseudo-labels better than confidence alone. Code is released, and the base comparison to VideoCutLER is reproduced with official checkpoints. Credit where due: the scaffolding is honest and the empirical work is substantial enough to be worth a referee's time.

The soft spot is the headline number. The 52.6 AP50 is obtained at a quality threshold τth=0.75, and Table 5 shows the result is very sensitive to that threshold: at 0.85 it drops to 48.8, at 0.95 to 48.7, at 0.50 to 52.4. The paper does not describe a held-out validation split; the threshold appears to have been selected directly on the YouTubeVIS-2019 val split, which is the same split used for the SOTA comparison. So the reported 4.4-point gain over VideoCutLER is optimistic, and at τth=0.85 the gain is only +0.6, within a plausible run-to-run noise range. The ablation attributing +2.1 to the quality predictor also compares against a confidence-only baseline tested at a single threshold (0.85) with no sweep, so that comparison is not fully controlled. No error bars or multiple seeds are reported for any of the main results. These are methodological holes, not necessarily fatal ones—the method's value does not entirely depend on the exact SOTA margin—but the paper should be revised to use a proper validation split or report all thresholds, and to tune the baselines fairly.

Who should read this? Researchers working on unsupervised video instance segmentation or on self-training with pseudo-labels will find the quality-guided filtering mechanism worth engaging with. The paper deserves a serious peer-review process, but it needs major revision on the evaluation protocol before the headline claim is credible.

Recommendation: send it to review, but ask for the threshold/validation issues to be resolved and for multi-run results.

Referee Report

4 major / 5 minor

Summary. The paper proposes AutoQ-VIS, an unsupervised video instance segmentation method. It starts from VideoCutLER's synthetic pseudo-labeled videos, jointly trains a VideoMask2Former instance segmentation model and a mask-quality predictor (Mask Scoring R-CNN style), and then performs multi-round self-training on unlabeled YouTubeVIS-2019 train videos. Pseudo-labels are filtered by a quality score Q_l = predicted IoU × confidence, and selected labels are fused into the training set; a DropLoss suppresses near-zero-overlap mask losses and model weights are reset each round. The headline result is 52.6 AP50 on the YouTubeVIS-2019 val split, reported as +4.4 AP50 over VideoCutLER, with ablations attributing gains to the quality predictor, DropLoss, and resetting.

Significance. The core idea—using a learned mask-IoU predictor rather than raw confidence to filter pseudo-labels in self-training—is timely and, if the result is robust, would be a solid step for annotation-free VIS. The paper reproduces several baselines from official code, releases source code, and explicitly compares against theoretical/practical upper bounds, which is commendable. The qualitative evidence and the Spearman correlation comparison (Q_l: 0.57 vs confidence: 0.42, Fig. 4) support the internal mechanism. However, the headline SOTA margin is currently tied to a quality threshold selected on the evaluation split, and the controlled comparison against the confidence baseline is incomplete. These issues must be resolved before the contribution can be considered fully established.

major comments (4)

[Table 5 / Sec. 3.2] The headline 52.6 AP50 is obtained at τth=0.75, but Table 5 shows strong sensitivity: τth=0.85 gives 48.8 AP50 and τth=0.95 gives 48.7 AP50, i.e., the gain over VideoCutLER (48.2) shrinks from +4.4 to +0.6. The paper does not mention a held-out validation split; τth appears to be selected directly on the YouTubeVIS-2019 val set, which is the same split used for the SOTA comparison. This makes the reported margin an optimistic estimate. Please report results with a validation split for threshold selection (e.g., a subset of the train set) or provide a principled, threshold-independent selection criterion.
[Table 3 and Sec. 3.2] The ablation 'w/o quality predictor' uses the confidence score s_l as the quality score with τth=0.85, while the full method uses Q_l with τth=0.75. Given the threshold sensitivity documented in Table 5, the +2.1 AP50 attributed to the quality predictor is not controlled: the baseline has not been swept over confidence thresholds, so it is unclear whether the gain comes from the predictor or from a more favorable operating point. Please sweep confidence thresholds for the baseline and report the best (or a matched-operating-point) comparison.
[Fig. 4 / Sec. 3.2] The quality predictor's transfer is load-bearing: it is trained only on synthetic pseudo-labels and then applied to real pseudo-labels in each self-training round, while the VIS model's weights drift. Fig. 4 shows a single scatter (Spearman 0.57 for Q_l vs 0.42 for confidence), but there are no error bars, multiple seeds, or per-round statistics. Since the quality predictor is frozen while the pseudo-label distribution changes across rounds, please report Spearman correlations and selection precision/recall for each self-training round, ideally with multiple seeds, to establish that ranking quality does not degrade.
[Tables 1, 3, 4, 5] No results include error bars or multiple seeds. Some differences in the paper are small (e.g., 52.6 vs 52.4 in Table 5; +0.6 AP50 at τth=0.85; '1 round' 51.3 vs '3 rounds' 52.0 in Table 4). Without variance estimates, it is difficult to assess which differences are meaningful. Please report at least 2–3 seeds for the main experiments and key ablations, or otherwise justify why single-run comparisons are sufficient in this setting.

minor comments (5)

[Abstract / Sec. 1] 'surpassing ... by 4.4%' is imprecise; the table reports 4.4 AP50 points. Please use '4.4 AP50' for consistency.
[Sec. 1] Typo: 'archives' should be 'achieves' in the contribution list.
[Table 5 caption] The caption text mentions 'τth = 0.80' but the table rows are 0.95, 0.85, 0.75, 0.50. Please correct the caption or the table.
[Sec. 5.2] Eq. (7) writes τ^(k) while Sec. 2.2 and Table 5 describe a fixed threshold τth. Clarify whether the threshold is fixed or round-dependent; if fixed, remove the superscript.
[Fig. 2] The text says 'four convolution layers' but the figure legend mentions kernel size and stride in a way that is easy to misread. Consider labeling the conv layer count directly on the figure.

Circularity Check

1 steps flagged · score 4.0 of 10

Main 52.6 AP50 SOTA claim is selected by tuning τth on YouTubeVIS-2019 val; the core self-training derivation is otherwise self-contained.

fitted input called prediction [Table 5 / Sec. 3.2 (Quality score threshold τth analysis); Abstract; Sec. 3 Datasets]
"Optimal performance (52.6 AP50) emerges at τth = 0.75, balancing valid sample retention and noise suppression."

The headline 52.6 AP50 is the value at τth=0.75 from Table 5, a threshold scan; Sec. 3 says evaluation is on YouTubeVIS-2019 val and no held-out split is mentioned. Thus τth is fitted to the target split, making the SOTA number the best val AP among thresholds rather than an independent prediction. The +2.1 AP50 quality-predictor gain is also uncontrolled: AutoQ-VIS uses τth=0.75 while the 'w/o quality predictor' baseline uses τth=0.85 (Table 3), so the comparison mixes threshold selection with the predictor.

full rationale

No self-citation, imported uniqueness theorem, or ansatz-smuggling is present; the quality predictor is trained on VideoCutLER's synthetic pseudo-labels (an external source), and the final evaluation is against human annotations on YouTubeVIS-2019 val, so the central self-training derivation is not circular by construction. The only load-bearing reduction is the threshold-selection issue: τth is chosen from Table 5 on the same val split that is then used for the SOTA comparison, making the 52.6 AP50 an optimistic selected maximum rather than a held-out prediction. This is a partial circularity in the evaluation protocol, but the method's components (quality predictor, DropLoss, resetting) are still evaluated externally at the chosen configuration, so the paper does not reduce entirely to a fit. Score 4 reflects this one fitted-input-called-prediction step while acknowledging the method's independent content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method is an empirical training pipeline. The most important free parameter is the quality threshold, chosen by validation performance. The axioms are standard assumptions about the quality of the bootstrapping data and the transferability of the quality score. No new physical entities are introduced.

free parameters (5)

quality threshold tau_th = 0.75 (selected on YouTubeVIS-2019 val)
Chosen from Table 5 as the value giving highest val AP50; the reported 52.6 depends on this choice.
number of self-training rounds = 2
Selected based on ablation (Table 4) as peak AP50; 3 rounds degrade by 0.6 AP50.
confidence threshold for pseudo-label generation = 0.25
Fixed in Eq. (2), not swept; determines the initial candidate pool.
DropLoss IoU threshold tau_IoU = 0.01
Low threshold to filter only near-zero overlap predictions (Sec 2.3).
balanced sampling probability = 0.5
Equal probability of sampling synthetic vs pseudo-labeled videos (Supplementary 5.3).

assumptions (4)

domain assumption VideoCutLER's synthetic video dataset and pretrained VideoMask2Former model are accepted as provided
The method bootstraps from VideoCutLER's synthetic pseudo-labels; if these are low-quality or biased, the entire pipeline inherits the bias (Sec 2.1).
domain assumption CutLER pseudo-labels on ImageNet serve as ground truth for training the quality predictor
The quality predictor is supervised with threshold-binarized IoU against these pseudo-labels; no real ground truth is used (Sec 2.1).
domain assumption YouTubeVIS-2019 train split is a valid unlabeled source for self-training and matches the val distribution
Self-training assumes pseudo-labels on the train split contain useful signal for the target benchmark (Sec 3).
domain assumption The quality predictor's ranking (Spearman 0.57) transfers across self-training rounds
The predictor is frozen while the VIS model updates; its reliability under feature drift is assumed, not measured (Sec 3.2, Fig 4).

how reviews work

0 comments

Cite this review

Pith. "Pith review of AutoQ-VIS: Improving Unsupervised Video Instance Segmentation via Automatic Quality Assessment." pith.science (2026). https://pith.science/paper/LEX3MMGX

@misc{pith2026250819808,
  author       = {Pith},
  title        = {Pith review of: AutoQ-VIS: Improving Unsupervised Video Instance Segmentation via Automatic Quality Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LEX3MMGX}},
  note         = {Machine review of arXiv:2508.19808}
}

abstract

Video Instance Segmentation (VIS) faces significant annotation challenges due to its dual requirements of pixel-level masks and temporal consistency labels. While recent unsupervised methods like VideoCutLER eliminate optical flow dependencies through synthetic data, they remain constrained by the synthetic-to-real domain gap. We present AutoQ-VIS, a novel unsupervised framework that bridges this gap through quality-guided self-training. Our approach establishes a closed-loop system between pseudo-label generation and automatic quality assessment, enabling progressive adaptation from synthetic to real videos. Experiments demonstrate state-of-the-art performance with 52.6 $\text{AP}_{50}$ on YouTubeVIS-2019 val set, surpassing the previous state-of-the-art VideoCutLER by 4.4$\%$, while requiring no human annotations. This demonstrates the viability of quality-aware self-training for unsupervised VIS. The source code of our method is available at https://github.com/wcbup/AutoQ-VIS.

Figures

Figures reproduced from arXiv: 2508.19808 by the authors.

**Figure 1.** AutoQ-VIS overview. In the initial training stage, both the VIS model and the mask quality predictor are trained on synthetic videos with pseudo annotations [15]. During the multiround self-training stage, the VIS model generates pseudo masks on unlabeled videos, which are then scored by the frozen quality predictor. Pseudo masks with high predicted quality are selected and added to the training set. The VIS model … view at source ↗

**Figure 2.** Network architecture of VideoMask2Former [2, 3] and Mask Quality Predictor. Our quality predictor integrates mask predictions and pixel decoder features following [9], employing a sequential architecture with four convolution layers (3×3 kernels, final layer stride of 2 for spatial reduction) followed by three fully-connected layers that ultimately produce mask IoU predictions. tively generate pseudo-labels on real… view at source ↗

**Figure 3.** The qualitative results on YouTubeVIS-2019 val split. AutoQ-VIS demonstrates superior instance discovery capabilities compared to VideoCutLER [15]: (1) Enhanced multiobject detection capacity, particularly for semantically distinct instances (e.g., person and bull in Column 2); (2) Improved segmentation fidelity through precise boundary delineation (e.g., the leopard in Column 3). (3) Better comprehensive insta… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Visualized comparison of quality score Ql and confidence score sl. Here, ρs denotes the Spearman’s rank correlation coefficient. Subplot (a) visualizes quality scores Ql and their ground truth IoU. Subplot (b) visualizes confidence scores sl and their ground truth Io…

**Figure 5.** Figure 5: Qualitative results of our VIS model on YouTubeVIS-2019 val split [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Qualitative results of our quality predictor on YouTubeVIS-2019 train split. The quality scores are shown in the center of each pseudo label [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

[1]

Dense unsupervised learning for video segmentation

Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Dense unsupervised learning for video segmentation. Ad- vances in Neural Information Processing Systems , 34: 25308–25319, 2021. 1

work page 2021
[2]

Mask2former for video instance segmentation

Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexan- der Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2, 3

arXiv 2021
[3]

Masked-attention mask transformer for universal image segmentation

Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexan- der Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 3

work page 2022
[4]

Guess What Moves: Unsupervised Video and Image Segmentation by Anticipat- ing Motion

Subhabrata Choudhury, Laurynas Karazija, Iro Laina, An- drea Vedaldi, and Christian Rupprecht. Guess What Moves: Unsupervised Video and Image Segmentation by Anticipat- ing Motion. In British Machine Vision Conference (BMVC),

work page
[5]

Imagenet: A large-scale hierarchical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 1, 2

work page 2009
[6]

Unsupervised ob- ject segmentation in video by efficient selection of highly probable positive features

Emanuela Haller and Marius Leordeanu. Unsupervised ob- ject segmentation in video by efficient selection of highly probable positive features. In Proceedings of the IEEE inter- national conference on computer vision , pages 5085–5093,

work page
[7]

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. 2

work page 2016
[8]

Vita: Video instance segmentation via object token association

Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. Advances in neural information processing systems, 35:23109–23120, 2022. 1

work page 2022

Show all 26 references

[9]

Mask scoring r-cnn

Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6409–6418, 2019. 2

2019
[10]

Bootstrapping objectness from videos by relaxed common fate and visual grouping

Long Lian, Zhirong Wu, and Stella X Yu. Bootstrapping objectness from videos by relaxed common fate and visual grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14582– 14591, 2023. 1

2023
[11]

Video instance segmentation tracking with a modified vae architecture

Chung-Ching Lin, Ying Hung, Rogerio Feris, and Linglin He. Video instance segmentation tracking with a modified vae architecture. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 13147–13157, 2020. 1

2020
[12]

Em-driven unsupervised learning for efficient motion seg- mentation

Etienne Meunier, Ana ¨ıs Badoual, and Patrick Bouthemy. Em-driven unsupervised learning for efficient motion seg- mentation. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 45(4):4462–4473, 2022. 1

2022
[13]

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
[14]

Cut and learn for unsupervised object detection and instance segmentation

Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3124– 3134, 2023. 1, 3

2023
[15]

Videocutler: Surprisingly simple un- supervised video instance segmentation

Xudong Wang, Ishan Misra, Ziyun Zeng, Rohit Girdhar, and Trevor Darrell. Videocutler: Surprisingly simple un- supervised video instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22755–22764, 2024. 1, 2, 3, 4

2024
[16]

End-to-end video instance segmentation with transformers

Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8741–8750, 2021. 1

2021
[17]

Seqformer: Sequential transformer for video instance segmentation

Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In European Conference on Computer Vision, pages 553–569. Springer, 2022. 1

2022
[18]

Segment- ing moving objects via an object-centric layered representa- tion

Junyu Xie, Weidi Xie, and Andrew Zisserman. Segment- ing moving objects via an object-centric layered representa- tion. Advances in neural information processing systems, 35: 28023–28036, 2022. 1, 3

2022
[19]

Self-supervised video object segmentation by motion grouping

Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 7177–7188,
[20]

Video instance seg- mentation

Linjie Yang, Yuchen Fan, and Ning Xu. Video instance seg- mentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5188–5197, 2019. 1, 2, 3

2019
[21]

Bdd100k: A diverse driving dataset for heterogeneous multitask learning

Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar- rell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition ...

2020
[22]

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,
[23]

Dvis: Decoupled video instance segmentation framework

Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. Dvis: Decoupled video instance segmentation framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 1282–1291, 2023. 1

2023
[24]

A survey on deep learning technique for video segmentation

Tianfei Zhou, Fatih Porikli, David J Crandall, Luc Van Gool, and Wenguan Wang. A survey on deep learning technique for video segmentation. IEEE transactions on pattern analysis and machine intelligence, 45(6):7099–7122, 2022. 1 AutoQ-VIS: Improving Unsupervised Video Instance ...

2022
[25]

Detailed methodology This section supplements what is not clearly stated in Sec. 2.2. 5.1. Automated pseudo-annotation with spatio- temporal NMS After the training of the VIS model, we use it to label the unlabeled videos. Let D = {di}N i=1 denote the initial detec- tion set p...
[26]

5 and Fig

Additional qualitative visualizations We provide additional qualitative results of our VIS model and quality predictor in Fig. 5 and Fig. 6. Time Figure 5. Qualitative results of our VIS model on YouTubeVIS-2019val split. Time Figure 6. Qualitative results of our quality predi...

Pith tools

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

[1] [1]

Dense unsupervised learning for video segmentation

Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Dense unsupervised learning for video segmentation. Ad- vances in Neural Information Processing Systems , 34: 25308–25319, 2021. 1

work page 2021

[2] [2]

Mask2former for video instance segmentation

Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexan- der Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2, 3

arXiv 2021

[3] [3]

Masked-attention mask transformer for universal image segmentation

Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexan- der Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 3

work page 2022

[4] [4]

Guess What Moves: Unsupervised Video and Image Segmentation by Anticipat- ing Motion

Subhabrata Choudhury, Laurynas Karazija, Iro Laina, An- drea Vedaldi, and Christian Rupprecht. Guess What Moves: Unsupervised Video and Image Segmentation by Anticipat- ing Motion. In British Machine Vision Conference (BMVC),

work page

[5] [5]

Imagenet: A large-scale hierarchical image database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 1, 2

work page 2009

[6] [6]

Unsupervised ob- ject segmentation in video by efficient selection of highly probable positive features

Emanuela Haller and Marius Leordeanu. Unsupervised ob- ject segmentation in video by efficient selection of highly probable positive features. In Proceedings of the IEEE inter- national conference on computer vision , pages 5085–5093,

work page

[7] [7]

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. 2

work page 2016

[8] [8]

Vita: Video instance segmentation via object token association

Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. Advances in neural information processing systems, 35:23109–23120, 2022. 1

work page 2022