Pith. sign in

REVIEW 3 major objections 6 minor 31 references

QueenVIS: Rethinking Image-Only Training for Video Instance Segmentation via Query Enrichment

T0 review · 3 major / 6 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read Image-only training can nearly match video-supervised VIS if object queries are enriched with appearance and spatial priors.

desk verdict Solid image-only VIS gains via training-only query enrichment; the AP story holds, but the "zero extra parameters" claim is contradicted by their own Table S2 and needs a fix. read the letter →

arxiv 2607.24598 v1 pith:66M4C2UQ submitted 2026-07-27 cs.CV

Arian Kheirandish , Fardin Ayar , Ehsan Javanmardi , Manabu Tsukada , Mahdi Javanmardi This is my paper
classification cs.CV
keywords videoinstancesegmentationimage-onlytrainingobjectqueriesMask2Formerqueryenrichmenttraining-freeassociationappearancedistillationcenterprediction
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 usually depends on costly multi-frame training with identity-consistent labels so models learn to keep the same object tagged across time. This paper argues that much of that temporal signal is already latent in ordinary single-frame features, and that the real bottleneck is the quality of the transformer object queries themselves: when they are trained only to separate objects inside one image, they drift apart across frames and association fails. QueenVIS keeps a Mask2Former-style model trained on independent frames, but adds two training-only heads that force each query to match its instance’s pooled backbone appearance descriptor and to predict the instance’s center. Both heads are thrown away at test time, so the network gains no parameters. Identity over time is then handled by a simple, training-free scheme that warm-starts queries from the previous frame and matches against a short memory bank. The result closes much of the gap to recent video-trained trackers on standard benchmarks, with the largest gains on longer sequences, without ever seeing a video clip during training.

What carries the argument

Query enrichment: two auxiliary training-only heads on Mask2Former queries—a feature-prediction loss that regresses each matched query toward its instance’s masked-pooled backbone descriptor, and a center-prediction loss that regresses normalized box center—discarded at inference, paired with confidence-guided content-query blending and a non-parametric memory bank for association.

What would settle it

On long, heavily occluded sequences, measure whether QueenVIS’s tracking-oracle gap stays near the oracle ceiling; if identity switches and the T-Oracle gap remain large relative to strong video-trained trackers, the claim that enriched image-only queries suffice for association fails.

Watch

Extended reading notes

Core claim

Object-query quality, not video-level supervision, is the binding constraint for image-only video instance segmentation. Distilling per-frame backbone appearance into each query and supervising an absolute center prior during single-frame training produces embeddings stable enough that training-free bipartite matching and a short memory bank can maintain identities, reaching competitive accuracy with video-supervised online methods while adding zero inference parameters.

Load-bearing premise

The appearance already present in single-frame backbone features, once copied into the queries along with a simple center prior, is enough to keep identities stable across frames without any video training—especially when objects heavily occlude one another.

Editorial extensions

If this is right

  • Competitive VIS is possible without identity-consistent video annotations or multi-frame training clips.
  • Gains concentrate on longer sequences, where per-frame query drift would otherwise accumulate.
  • Association, not mask quality alone, accounts for most of the measured improvement under oracle diagnosis.
  • Future image-only VIS work can target better auxiliary objectives rather than more elaborate video trackers when labels or compute are scarce.

Reading between the lines

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

  • The same query-enrichment idea may transfer to other query-based video tasks (tracking, panoptic, referring segmentation) that currently assume clip-level training.
  • Dense occlusion remains the regime where pure image-only priors are most likely to need complementary cues such as motion or depth.
  • If classification residual, not association, becomes the next bottleneck after enrichment, redesigning the class head may yield further image-only gains.
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 manuscript presents QueenVIS, an image-only-trained video instance segmentation framework built on Mask2Former/MinVIS. The core idea is to enrich object queries during single-frame training with two auxiliary, training-only losses: a feature-prediction loss that regresses each matched query toward the masked-average-pooled backbone descriptor (Res4+Res5) of its instance, and a center-prediction loss that regresses each query toward its instance's normalized bounding-box center. Both heads are claimed to be discarded at inference. Temporal association is handled by a training-free confidence-guided query-propagation rule (Eq. 5) plus a non-parametric memory bank adapted from VISAGE. With a ResNet-50 backbone the method reports 50.9 AP on YouTube-VIS 2021 (+6.7 over MinVIS), 51.8 on 2019 (+4.4), 29.8 on OVIS (+4.8), and 33.6 on the YouTube-VIS 2022 long split (+10.3). A diagnostic study using the T-/TC-Oracle framework of Hamdi et al. attributes the gains almost entirely to improved association (T-Oracle gap 12.5→7.5 AP; AR1 tracking gap to 1.0). Ablations cover component increments (Table 2), feature levels (Table 5), inference hyperparameters (Table 6), loss-weight sensitivity (Table S4), relational-vs-absolute distillation (Table S3), and Swin-L scaling (Table S1). The paper is clearly written, the evaluation is broad, and the honest reporting of the remaining OVIS gap is commendable.

Significance. If the results hold, the paper makes a useful and somewhat contrarian point: much of what video-supervised VIS methods gain from temporal supervision can be recovered by better shaping the per-frame query representation, at no inference cost and without identity-consistent video annotations. Strengths that weigh positively: (i) multi-benchmark evaluation on four public benchmarks against external published numbers, so there is no evaluation circularity — auxiliary targets come from backbone features and GT image masks, not from the AP metric; (ii) an incremental ablation (Table 2) separating inference-only gains (+2.2) from training gains (+4.5); (iii) a model-agnostic oracle analysis (Table 4) that localizes the improvement to association rather than mask/classification quality — a genuinely falsifiable attribution; (iv) sensitivity sweeps showing the method is not a tuned hyperparameter point (Tables 6, S4); (v) a negative result (relational distillation, Table S3) reported with a mechanistic explanation; (vi) promised code release. The work would be a solid reference point for the image-only VIS line. Two attribution issues (detailed below) currently prevent full confidence in *

major comments (3)
  1. [§3.4, §B, Table S2] The 'zero additional parameters at inference' claim — made in the abstract, §3.4, and §B — is contradicted by the paper's own Table S2, which reports 300.34M parameters for QueenVIS vs 293.89M for MinVIS (a +6.45M difference) at identical GFLOPs (266.09). Identical GFLOPs suggest the extra parameters are not exercised, which makes them harder to account for. Two readings have different consequences: (a) the discarded aux heads were retained in the inference checkpoint — then the headline efficiency claim is simply false as written and must be corrected; (b) the extra parameters come from the VISAGE recipe the authors say they follow 'unchanged except that we remove its video-level supervision' (§4.1) — then the inference network is not Mask2Former-plus-nothing, and the ablation attribution in §4.5 is confounded by capacity/modules the ablations do not isolate. The authors must identify t
  2. [§4.1, §4.5, Table 2] The attribution of the +4.5 AP training gain to query enrichment rests on a missing control. Per §4.1, QueenVIS is trained with the VISAGE recipe (schedule, augmentations, hyperparameters), whereas the ablation baseline rows in Table 2 use the official MinVIS weights trained with the MinVIS recipe. The jump from 46.4 (MinVIS official + QP + MB) to 49.6 (+ FP) therefore conflates the feature-prediction loss with the change of training recipe. The paper asserts the recipe choice ensures 'any gain is attributable to query enrichment rather than to retuning the baseline,' but this is only true if a Mask2Former trained under the VISAGE recipe *without* the auxiliary losses (evaluated with QP + MB) is shown to match the official MinVIS baseline. This control requires a single additional training run and is load-bearing for the central claim that 'the larger improvement comes from training rath
  3. [§3.5, Eq. (5)] The query-propagation rule is under-specified in a way that affects reproducibility. Eq. (5) blends C_i^{t-1} into 'the same query' slot i at frame t, but Mask2Former content queries are slot-indexed and instance-to-slot assignment within a frame is permutation-dependent; identity is only established *after* the decoder via Hungarian matching against the memory bank. What is the cross-frame correspondence between slot i at t-1 and slot i at t at propagation time — raw slot index, or a matching-based remap? If raw index, the propagated warm start will frequently inject the wrong instance's embedding after slot permutations, and the mechanism by which Eq. (5) still helps needs explanation. Please specify the correspondence rule and, ideally, ablate index-based vs matching-based propagation.
minor comments (6)
  1. [Figure 1] The t-SNE figure is presented as diagnostic evidence for query drift, but the caption states the red ovals 'explicitly denote instances where QueenVIS successfully maintains tight clusters,' i.e., the highlighted cases are selected. Please describe the sampling/matching protocol (how many identities, how queries were paired across the two models) and avoid selection-based highlighting, or move the figure to qualitative status.
  2. [Table 3] The YouTube-VIS 2022 long-split comparison includes only MinVIS. Since +10.3 AP on long sequences is a headline claim ('query quality, rather than video-level supervision, is the binding constraint'), including published video-supervised numbers on this split would contextualize how much of the long-video gap actually closes.
  3. [§4.5 / §B] 'Throughput drop of less than 7 FPS' understates the cost: 26.92→20.02 FPS is a ~26% relative reduction, and 20 FPS is marginal for some real-time applications. Please report the relative overhead and the memory-bank matching cost separately.
  4. [§3.3] c*_i is defined as the bounding-box center; for elongated or articulated masks the mask centroid may differ substantially. A one-line justification (or a centroid ablation) would help. Also state whether center prediction uses normalized coordinates per frame and how it interacts with large camera motion.
  5. [§4.4, Table 4] The oracle framework of Hamdi et al. [12] is an arXiv preprint from the same group. The analysis is valuable, but please confirm the oracle implementation and metric definitions are public/reproducible, and soften 'tighter than even video-supervised trackers achieve' to reflect that the comparison pool in Table 4 is small (three methods).
  6. [Abstract/§1] '+6.7 AP on YouTube-VIS' should specify the 2021 split; the 2019 gain is +4.4. Minor typographical issues: 'QueenVIS MinVIS' appears as a stray caption fragment in Figure 1; spacing artifacts ('training-freeandfeeds', 'instanceswithina') suggest a LaTeX issue.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity: empirical VIS gains on public benchmarks; mild non-load-bearing self-citation of authors' oracle/panoptic tools.

  1. self citation load bearing [§4.4 Table 4; Ref. [12]]
    "Using the model-agnostic oracle framework of Hamdi et al. [12], we attribute this gain to association, shrinking the tracking-recall gap to essentially the oracle ceiling... Table 4: Diagnostic error gap analysis on YouTube-VIS 2021 using the framework of Hamdi et al. [12]."

    The diagnostic that attributes gains to association relies on an arXiv oracle framework by overlapping authors (Ayar, Javanmardi). This is mild self-citation used for interpretation, not for constructing the main AP results; the primary claims still rest on public VIS metrics. Not load-bearing for the derivation of the method or the reported AP numbers.

full rationale

QueenVIS's central claim is empirical: single-frame training with two auxiliary heads (feature regression to stop-grad backbone pooled descriptors; center regression to GT box centers) plus training-free association improves AP over MinVIS on YouTube-VIS/OVIS. Those AP numbers are external benchmark metrics, not quantities defined by the auxiliary losses or by fitted identity labels. L_feat and L_center supervise query embeddings toward fixed per-frame targets; they do not construct video AP or force bipartite matching outcomes by definition. Inference association is explicitly training-free and not optimized on video identity supervision. Self-citation appears only as methodological tooling—the Hamdi et al. oracle diagnostic [12] (overlapping authors) and the center-prediction idea from the authors' panoptic work [11]—neither of which is a uniqueness theorem or a fitted input renamed as a prediction. Parameter-count inconsistency in Table S2 is a correctness/bookkeeping issue, not circular derivation. Overall the chain is self-contained against external benchmarks.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The work is empirical deep learning on public VIS benchmarks. It inherits Mask2Former/MinVIS set-prediction and Hungarian matching, assumes backbone masked-pooled features are good appearance teachers, and treats a short non-parametric memory plus confidence blending as enough temporal glue. Free parameters are standard loss and inference knobs selected on validation. No new physical entities are postulated.

free parameters (6)
  • λ_feat (feature-prediction loss weight) = 0.2
    Balances appearance regression against classification/mask losses; swept in Table S4 and fixed at 0.2 for main results.
  • λ_center (center-prediction loss weight) = 0.05
    Weights spatial center regression; swept and fixed at 0.05 for the final model.
  • α (query-propagation blend factor) = 0.25
    Controls how much previous-frame content query is mixed into the current init; default 0.25 after sensitivity sweep (Table 6).
  • τ (propagation confidence threshold) = 0.8
    Gates which previous queries are propagated; default 0.8 from Table 6 sweep.
  • K (memory bank length) = 3
    Number of recent frames aggregated for matching; default 3 from ablation.
  • Feature levels for teacher descriptor (Res4+Res5) = Res4 + Res5 concatenation
    Which backbone stages are pooled and concatenated as the appearance target; chosen by Table 5 validation AP.
assumptions (5)
  • domain assumption Mask2Former-style object queries optimized with classification and mask losses, plus Hungarian matching within a frame, form a valid base for instance segmentation.
    Entire architecture is built on Mask2Former/MinVIS without re-deriving set prediction (§2.1, §3.1).
  • domain assumption Masked-average-pooled backbone features of a ground-truth instance are a sufficiently discriminative appearance teacher for the matched query.
    Core of L_feat in §3.2 (Eqs. 1–2); stop-grad on the teacher assumes backbone descriptors already carry the needed signal.
  • domain assumption Absolute normalized box center is a useful spatial prior that reduces long-range identity switches when appearance is ambiguous.
    Motivated in §3.3 and tied to prior video panoptic work; supervised by L_center (Eq. 3).
  • domain assumption Training-free bipartite matching on enriched queries, with confidence-guided content blending and a short non-parametric memory, can maintain identities without video-supervised track heads.
    Inference pipeline §3.5; memory bank adapted from VISAGE but fed image-only queries.
  • standard math Standard ℓ1/ℓ2-normalized regression and Dice/CE/mask losses are appropriate supervision for the stated heads.
    Loss forms in Eqs. 2–4 follow common detection/segmentation practice.
invented entities (2)
  • QueenVIS feature-prediction and center-prediction auxiliary heads (training-only)
    purpose: Distill backbone appearance and explicit center geometry into object queries under image-only supervision, then discard at inference.
    Not new physical entities; engineering modules. Independent evidence is indirect via AP/oracle gains and ablations rather than an external predicted observable.
  • Confidence-guided content-query propagation rule (Eq. 5)
    purpose: Warm-start current-frame content queries from high-confidence previous embeddings without a learned tracker.
    Inference heuristic specific to this pipeline; validated only by ablations on VIS benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QueenVIS: Rethinking Image-Only Training for Video Instance Segmentation via Query Enrichment." pith.science (2026). https://pith.science/paper/66M4C2UQ

@misc{pith2026260724598,
  author       = {Pith},
  title        = {Pith review of: QueenVIS: Rethinking Image-Only Training for Video Instance Segmentation via Query Enrichment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66M4C2UQ}},
  note         = {Machine review of arXiv:2607.24598}
}
read the original abstract

Video instance segmentation (VIS) requires models to detect, segment, and track object identities across frames, and most methods enforce temporal consistency through video-level supervision. Image-only training approaches, with MinVIS as one prominent example, have challenged this assumption, reaching competitive VIS without video training by treating frames as independent images and associating instances only at inference. The field has nonetheless moved toward ever more elaborate video-trained trackers, which depend on costly identity-consistent annotations, leaving the image-only direction under-explored. A diagnostic analysis identifies object query quality as the bottleneck: queries trained only to localize objects within a frame drift apart across frames and destabilize association. QueenVIS introduces a query-centric framework for strengthening image-trained VIS. During single-frame training, we enrich Mask2Former queries with two auxiliary heads: a feature-prediction loss that aligns each query with the pooled backbone descriptor of its instance, and a center-prediction loss that injects spatial structure. Both heads are discarded at inference, adding zero parameters, and temporal identity is maintained by a training-free query-propagation and memory-bank scheme. On YouTube-VIS and OVIS with a ResNet-50 backbone, QueenVIS improves over MinVIS, up to +6.7 AP on YouTube-VIS, +4.8 AP on OVIS, and +10.3 AP on the long-sequence YouTube-VIS split. QueenVIS achieves 50.9 AP on YouTube-VIS and remains competitive with recent video-supervised state-of-the-art, without processing a single video clip during training. Our findings suggest that strengthening the discriminative power and temporal stability of object queries is an important, underexplored axis for VIS. Code and models: https://github.com/ArianKheir/QueenVIS

Figures

Figures reproduced from arXiv: 2607.24598 by the authors.

Figure 1
Figure 1. t-SNE visualization of object queries on the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed appearance-aware query learning framework. To preserve the efficiency of image-only optimization, the architecture is divided into two distinct phases. (a) Image-Only Training Pipeline: We introduce auxiliary feature and center prediction branches (active only during training). Ground-truth features are extracted by pooling backbone feature maps using target masks, supervising the transforme… view at source ↗
Figure 3
Figure 3. Qualitative comparison on the OVIS dataset. The rows correspond to the original input frames (top), MinVIS [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 3 linked inside Pith

  1. [1]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kir- illov, and S. Zagoruyko, “End-to-end object detection with transformers,” inEuropean conference on computer vision, pp. 213–229, Springer, 2020

    2020

  2. [2]

    Masked-attention mask transformer for universal image segmentation,

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-attention mask transformer for universal image segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1290–1299, 2022

    2022

  3. [3]

    Seq- former: Sequential transformer for video instance seg- mentation,

    J. Wu, Y . Jiang, S. Bai, W. Zhang, and X. Bai, “Seq- former: Sequential transformer for video instance seg- mentation,” inEuropean Conference on Computer Vi- sion, pp. 553–569, Springer, 2022

    2022

  4. [4]

    In defense of online models for video instance segmen- tation,

    J. Wu, Q. Liu, Y . Jiang, S. Bai, A. Yuille, and X. Bai, “In defense of online models for video instance segmen- tation,” inEuropean Conference on Computer Vision, pp. 588–605, Springer, 2022

    2022

  5. [5]

    Visage: Video instance segmentation with appearance-guided enhancement,

    H. Kim, J. Kang, M. Heo, S. Hwang, S. W. Oh, and S. J. Kim, “Visage: Video instance segmentation with appearance-guided enhancement,” inEuropean Confer- ence on Computer Vision, pp. 93–109, Springer, 2024

    2024

  6. [6]

    Minvis: A minimal video instance segmentation framework without video-based training,

    D.-A. Huang, Z. Yu, and A. Anandkumar, “Minvis: A minimal video instance segmentation framework without video-based training,”Advances in Neural Information Processing Systems, vol. 35, pp. 31265–31277, 2022

    2022

  7. [7]

    A generalized framework for video instance segmentation,

    M. Heo, S. Hwang, J. Hyun, H. Kim, S. W. Oh, J.-Y . Lee, and S. J. Kim, “A generalized framework for video instance segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14623–14632, 2023

    2023

  8. [8]

    Tcovis: Tempo- rally consistent online video instance segmentation,

    J. Li, B. Yu, Y . Rao, J. Zhou, and J. Lu, “Tcovis: Tempo- rally consistent online video instance segmentation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 1097–1107, 2023

    2023

Show all 31 references
  1. [9]

    Dvis: Decoupled video instance segmentation framework,

    T. Zhang, X. Tian, Y . Wu, S. Ji, X. Wang, Y . Zhang, and P. Wan, “Dvis: Decoupled video instance segmentation framework,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pp. 1282–1291, 2023

    2023

  2. [10]

    Cavis: Context-aware video instance segmentation,

    S. Lee, J. Seo, K. Han, M. Choi, and S. Im, “Cavis: Context-aware video instance segmentation,” inPro- ceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4507–4517, 2025

    2025

  3. [11]

    Lidar-camera fusion for video panoptic segmentation without video training,

    F. Ayar, E. Javanmardi, M. Tsukada, M. Javanmardi, and M. Rahmati, “Lidar-camera fusion for video panoptic segmentation without video training,” inProceedings of the 2024 International Conference on Intelligent Com- puting and Its Emerging Applications (ICEA), pp. 1–7, 2024

    2024

  4. [12]

    Mind the gap: Disentangling performance bottlenecks in video instance segmentation,

    D. Hamdi, F. Ayar, and M. Javanmardi, “Mind the gap: Disentangling performance bottlenecks in video instance segmentation,”arXiv preprint arXiv:2606.07394, 2026

    2026 arXiv

  5. [13]

    Per-pixel classification is not all you need for semantic segmen- tation,

    B. Cheng, A. G. Schwing, and A. Kirillov, “Per-pixel classification is not all you need for semantic segmen- tation,” inAdvances in Neural Information Processing Systems, vol. 34, pp. 17864–17875, 2021

    2021

  6. [14]

    Video instance segmen- tation,

    L. Yang, Y . Fan, and N. Xu, “Video instance segmen- tation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 5188–5197, 2019

    2019

  7. [15]

    Crossover learning for fast on- line video instance segmentation,

    S. Yang, Y . Fang, X. Wang, Y . Li, C. Fang, Y . Shan, B. Feng, and W. Liu, “Crossover learning for fast on- line video instance segmentation,” inProceedings of the IEEE/CVF International Conference on Computer Vi- sion (ICCV), pp. 8043–8052, October 2021

    2021

  8. [16]

    VidEoMT: Your ViT is secretly also a video segmentation model,

    N. Norouzi, I. E. Zulfikar, N. Cavagnero, T. Kerssies, B. Leibe, G. Dubbelman, and D. de Geus, “VidEoMT: Your ViT is secretly also a video segmentation model,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 35177– 35186, 2026. 9

    2026

  9. [17]

    NOVIS: A case for end-to-end near- online video instance segmentation,

    T. Meinhardt, M. Feiszli, Y . Fan, L. Leal-Taix ´e, and R. Ranjan, “NOVIS: A case for end-to-end near- online video instance segmentation,”arXiv preprint arXiv:2308.15266, 2023

    2023 arXiv

  10. [18]

    Efficient video instance segmentation via tracklet query and proposal,

    J. Wu, S. Yarram, H. Liang, T. Lan, J. Yuan, J. Eledath, and G. Medioni, “Efficient video instance segmentation via tracklet query and proposal,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 959–968, 2022

    2022

  11. [19]

    Ctvis: Con- sistent training for online video instance segmentation,

    K. Ying, Q. Zhong, W. Mao, Z. Wang, H. Chen, L. Y . Wu, Y . Liu, C. Fan, Y . Zhuge, and C. Shen, “Ctvis: Con- sistent training for online video instance segmentation,” inProceedings of the IEEE/CVF International Confer- ence on Computer Vision, pp. 899–908, 2023

    2023

  12. [20]

    Visolo: Grid-based space-time aggregation for efficient online video instance segmen- tation,

    S. H. Han, S. Hwang, S. W. Oh, Y . Park, H. Kim, M.- J. Kim, and S. J. Kim, “Visolo: Grid-based space-time aggregation for efficient online video instance segmen- tation,” inproceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 2896–2905, 2022

    2022

  13. [21]

    Video instance segmentation using inter-frame communication transformers,

    S. Hwang, M. Heo, S. W. Oh, and S. J. Kim, “Video instance segmentation using inter-frame communication transformers,”Advances in Neural Information Process- ing Systems, vol. 34, pp. 13352–13363, 2021

    2021

  14. [22]

    Vita: Video instance segmentation via object token as- sociation,

    M. Heo, S. Hwang, S. W. Oh, J.-Y . Lee, and S. J. Kim, “Vita: Video instance segmentation via object token as- sociation,”Advances in neural information processing systems, vol. 35, pp. 23109–23120, 2022

    2022

  15. [23]

    Mdqe: Mining discriminative query embeddings to segment occluded instances on challenging videos,

    M. Li, S. Li, W. Xiang, and L. Zhang, “Mdqe: Mining discriminative query embeddings to segment occluded instances on challenging videos,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10524–10533, 2023

    2023

  16. [24]

    Re- finevis: Video instance segmentation with temporal at- tention refinement,

    A. Abrantes, J. Wang, P. Chu, Q. You, and Z. Liu, “Re- finevis: Video instance segmentation with temporal at- tention refinement,”arXiv preprint arXiv:2306.04774, 2023

    2023 arXiv

  17. [25]

    Sipmask: Spatial information preserva- tion for fast image and video instance segmentation,

    J. Cao, R. M. Anwer, H. Cholakkal, F. S. Khan, Y . Pang, and L. Shao, “Sipmask: Spatial information preserva- tion for fast image and video instance segmentation,” inEuropean Conference on Computer Vision, pp. 1–18, Springer, 2020

    2020

  18. [26]

    The 3rd large-scale video object segmentation challenge – video instance segmentation track,

    L. Yang, Y . Fan, Y . Fu, and N. Xu, “The 3rd large-scale video object segmentation challenge – video instance segmentation track,” June 2021

    2021

  19. [27]

    The 4th large-scale video object segmentation challenge – video instance segmen- tation track,

    L. Yang, Y . Fan, and N. Xu, “The 4th large-scale video object segmentation challenge – video instance segmen- tation track,” June 2022

    2022

  20. [28]

    Occluded video instance segmentation: A benchmark,

    J. Qi, Y . Gao, Y . Hu, X. Wang, X. Liu, X. Bai, S. Be- longie, A. Yuille, P. H. S. Torr, and S. Bai, “Occluded video instance segmentation: A benchmark,”Interna- tional Journal of Computer Vision, vol. 130, no. 8, pp. 2022–2039, 2022

    2022

  21. [29]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recog- nition, pp. 770–778, 2016

    2016

  22. [30]

    Relational knowledge distillation,

    W. Park, D. Kim, Y . Lu, and M. Cho, “Relational knowledge distillation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 3967–3976, 2019

    2019

  23. [31]

    Swin Transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin Transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 10012–10022, 2021. 10 Supplementary Material...

    2021

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.