Pith. sign in

REVIEW 3 major objections 5 minor 95 references

Rethinking Query-based Transformer for Continual Image Segmentation

T0 review · 3 major / 5 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 SimCIS claims that object queries pre-aligned to image features preserve built-in objectness across continual stages, yielding state-of-the-art ADE20K panoptic and semantic segmentation while remaining stable across class orders and using…

desk verdict SimCIS's query pre-alignment is a genuinely useful new mechanism, but the paper's across-setting claims outrun the evidence: only the overlap protocol is evaluated, and the 10x storage claim doesn't match the tables. read the letter →

arxiv 2507.07831 v1 pith:3Q7AQEXM submitted 2025-07-10 cs.CV

Yuchen Zhu , Cheng Shi , Dingyou Wang , Jiajin Tang , Zhengxuan Wei , Yu Wu , Guanbin Li , Sibei Yang This is my paper · ORCID
classification cs.CV
keywords continualimagesegmentationclass-incrementallearningquery-basedtransformerobjectnesscatastrophicforgettingvirtualqueryreplaypanopticsemantic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that the built-in objectness of query-based segmentation transformers—the property that keeps mask proposals alive during continual training—is really an alignment between object queries and semantic clusters already present in the image feature map, and that this alignment decays across stages. On that diagnosis it builds SimCIS, a continual segmentation method that pre-aligns queries by selecting the most prototype-similar image features as query vectors, adds a cross-stage consistency loss to keep those selections stable, and replays stored query vectors instead of images to prevent class forgetting. The reported experiments on ADE20K continual panoptic and semantic segmentation show consistent gains over prior methods, including under shuffled and descending class orders, with replay storage reduced roughly tenfold. A caveat is that the consistency loss is only computable when the same image appears in consecutive stages, so the method is tested in the overlap protocol rather than the standard disjoint protocol.

What carries the argument

The load-bearing mechanism is 'lazy Query Pre-Alignment' (QPA): instead of learning N object queries from scratch, the model selects N pixel-feature locations with the highest dot-product similarity to class prototypes and sets the queries to those feature vectors, with stop-gradient protecting the feature map. 'Consistent Selection Loss' (CSL) then uses a KL divergence between similarity distributions to make the selected locations for a given image match the previous stage's selection. 'Virtual Query' (VQ) stores matched decoder queries in a class-wise bank and replays them into the FFN layers only, skipping attention, to preserve class semantics without storing images. Together these turn built-in objectness from a property the model happens to have into an explicit assignment mechanism.

What would settle it

Train SimCIS on disjoint ADE20K 100-5 and 100-10 splits where each image is annotated once and Eq. 6 cannot be computed, then compare the full method against an ablation without CSL under several random class orders; if the CSL contribution nearly vanishes or the variance across orders matches the frozen baselines, the paper's cross-stage objectness and order-stability claims would not transfer.

Watch

Extended reading notes

Core claim

The paper argues that built-in objectness in Mask2Former arises because object queries learn to align with semantic priors already present in the pixel feature map, and that this alignment is what protects mask proposals from catastrophic forgetting. It then claims that this objectness decays across continual stages because background semantic shift and newly introduced classes break the query-feature alignment. SimCIS restores the alignment by lazy Query Pre-Alignment: at each stage, the model selects the N pixel-feature locations with the highest similarity to trainable class prototypes and uses those features as the object queries, with stop-gradient applied so the feature map stays stable. A Consistent Selection Loss forces the selected locations for the same image to match those chosen in the previous stage, and a Virtual Query bank replays class-matched decoder query vectors through the decoder's FFN layers only, bypassing attention, to combat class forgetting without storing images. The paper reports consistent improvements over prior methods on continual panoptic and semantic segmentation, including near-joint performance in short task sequences and stable results under shuffled input orders.

Load-bearing premise

The Consistent Selection Loss (Eq. 6) assumes the same image has been processed at the previous stage, so the previously selected feature positions still exist for the current image; all experiments use the overlap protocol where images repeat across tasks, and the method is never tested in the standard disjoint protocol where each image is annotated only once.

Editorial extensions

If this is right

  • Decoupling mask generation by freezing parameters is unnecessary; pre-aligning queries from image features preserves objectness while keeping training open for new classes.
  • New-class plasticity is maintained even in short task sequences where frozen-mask methods collapse to or below the fine-tuning baseline.
  • Class replay can be done with stored query vectors instead of images, cutting disk usage to roughly a tenth at comparable or better PQ while avoiding image-specific order effects.
  • The model stays within a few PQ or mIoU points of the joint oracle on short sequences, and on 100-50 panoptic segmentation it matches or exceeds the joint oracle on base classes.
  • Performance under shuffled and even descending class orders remains stable, unlike the frozen-prompt baseline whose all-class PQ drops sharply when classes are learned in descending order.

Reading between the lines

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

  • Because Eq. 6 needs the previous stage's feature map for the same image, the CSL component is validated only in the overlap protocol; a reader should not assume QPA and VQ alone reproduce the reported gains on disjoint splits.
  • The same pre-alignment recipe could be dropped into other Mask2Former-style universal segmenters, which the paper lists as future work; a direct test would be whether prototype-based query selection transfers without the CSL term.
  • The storage and privacy advantages of VQ depend on query vectors encoding category semantics; whether stored queries transfer across datasets or generalize to unseen object shapes is untested.
  • The descending-order experiment shows the method is stable when rare classes are learned first, but the ten random orders are permutations of ADE20K's fixed class ranking; robustness to arbitrary image-level or intra-stage ordering is a separate question.
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 / 5 minor

Summary. The paper proposes SimCIS, a continual image segmentation method built on Mask2Former, with three components: Lazy Query Pre-alignment (QPA) initializes object queries by selecting top-k pixel features with highest similarity to per-class prototypes; Consistent Selection Loss (CSL) enforces stability of selected feature locations across stages via a KL divergence between similarity distributions; Virtual Query (VQ) stores matched decoder queries per class and replays them through a skip-attention path to preserve class predictions. Experiments on ADE20K for continual panoptic and semantic segmentation report state-of-the-art results across multiple task splits (Tables 1-3), component ablations (Table 4), a storage-efficiency comparison (Table 5), and ten random class-order runs (Table 7). The method's core idea is to directly select image features for query assignment, preserving objectness while allowing new-class plasticity.

Significance. If the reported results hold, SimCIS is a simple and effective baseline that avoids freezing mask-generation parameters, addresses the plasticity and input-order sensitivity issues of decoupled frameworks, and reduces replay storage by about tenfold compared with image replay. The paper's strengths include extensive comparisons across panoptic and semantic segmentation settings, a clear component-wise ablation, a thoughtful storage-efficiency analysis, and a novel random-order evaluation with ten orders. However, the evaluation is restricted to the overlap protocol, the random-order experiments only compare against ECLIPSE, and no variance estimates or multiple-seed runs are reported; these gaps leave the claimed across-setting and across-order generality partially unsupported.

major comments (3)
  1. [Sec. 4.2, Eq. (6); Supplementary Sec. 8] The Consistent Selection Loss is motivated by keeping, "for the same image, the most semantically significant locations selected in the previous stage" (Sec. 4.2). In the disjoint protocol, each image is annotated only once, so no image is revisited across stages; although Eq. (6) is technically computable by running the frozen stage-(t-1) model on the current image, its semantic role degenerates from cross-stage selection consistency on the same image to a consistency regularizer between the frozen and current models on a new image. The manuscript states in Supplementary Sec. 8 that only the overlap protocol is used, and Tables 1-3 therefore contain no disjoint-protocol results. This leaves the claimed "across settings" generality unsupported exactly where the core novelty is stressed; please add disjoint-protocol results for the main tables and an ablation of CSL without overlap, or explicitly restrict the claims to the overlap setting.
  2. [Sec. 5.4, Table 7] The robustness-to-input-order claim is supported only by random-order experiments against ECLIPSE, only under the overlap protocol, and only for the 100-5 panoptic segmentation scenario. No random-order comparison is made with BalConpas or other state-of-the-art methods, no random-order results are reported for semantic segmentation or for the 50-base-class settings, and no disjoint-protocol random-order experiments are provided. Because the abstract and contribution list claim consistency "across ... input data orders," this evidence base is too narrow to support the claim; please broaden the comparison and report variance across the random orders.
  3. [Tables 1-3, 7] No error bars or multiple-seed runs are reported anywhere in the experimental section. Continual learning results are known to be sensitive to initialization, hyperparameters, and data ordering, so single-run numbers make the "consistently outperforms" claim fragile. Please report at least three seeds with standard deviations for the main tables and for the random-order table, or justify why a single run is sufficient for the conclusions.
minor comments (5)
  1. [Abstract] The abstract promises that "All models and codes will be made publicly available" with a GitHub link, but no code or model is provided with the submission; please clarify availability or provide the link.
  2. [Sec. 9, Table 7] The sentence "our model achieves an increase of 41.9% across all classes compared to ECLIPSE" is ambiguous: it likely refers to a relative improvement in average PQ, but the text reads as an absolute increase. Please state the comparison metric explicitly.
  3. [References] References [2] and [9] are the same paper (Baek et al., Decomposed Knowledge Distillation), and references [58] and [59] are also duplicates (Rusu et al., Progressive Neural Networks); please remove the duplicates.
  4. [Sec. 4.3, Eq. (9)] The exponent 1/2 in the pseudo-distribution weighting formula (Eq. 9) is introduced without justification or sensitivity analysis; please state how this value was chosen and whether results are robust to it.
  5. [Table 5] The header "Reply Num Disk" appears to be a typo; it should be "Replay Number Disk" or similar.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SimCIS's components are explicit architectural and loss choices, and the reported benchmark gains are external empirical comparisons rather than predictions derived from fitted inputs.

full rationale

SimCIS is an empirical method paper rather than a derivation chain, so the circularity patterns do not arise. The closest step to 'by construction' is Query Pre-Alignment: Eq. 4 selects top-K feature locations by prototype similarity and Eq. 5 initializes QN from those locations, so the query and the selected feature are identical by definition; however, the paper presents this as an explicit design to enforce alignment, not as a prediction or discovered result. Its effectiveness is then tested on held-out benchmarks (Tables 1-3) against external methods, which is independent evidence. The Consistent Selection Loss (Eq. 6) is a training regularizer, not a fitted quantity later renamed as a prediction; the CSL weight (2.0) and queue length h are hand-chosen hyperparameters. Virtual Query (Eqs. 7-11) is a replay mechanism validated by ablations and memory comparisons. Self-citations in the reference list (e.g., [38], [65], [69]) appear in related-work contexts and are not load-bearing for the central claim. A genuine limitation, but not a circularity, is that the random-order evidence (Table 7 and Table 6) only compares against ECLIPSE under the overlap protocol, and Supplementary Sec. 8 states 'We choose the more challenging overlap setting as our experimental protocol' without disjoint-protocol results; this weakens the broad claim of robustness across all settings and orders, yet it does not make any result equivalent to its own input. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The method depends on a handful of hyperparameters and on the overlap-protocol assumption. No numerically fitted constants are dressed up as predictions.

free parameters (4)
  • CSL loss weight = 2.0
    Set following Mask2Former defaults for other losses; no ablation shown for this weight (Sec. 8).
  • Number of virtual queries sampled per iteration (j) = up to 80
    The paper says 'N be set up to 80' in Sec. 5.1, but the actual j per batch is not specified.
  • Virtual query queue length h = not specified
    A queue of length h per class is introduced in Eq. 7; h is never given, nor is its effect ablated separately (only total memory size is reported).
  • Pseudo-distribution exponent = 1/2
    Eq. 9 uses exponent 1/2 in omega computation; no derivation or ablation for this choice.
assumptions (4)
  • domain assumption Pixel-decoder features F remain semantically clustered across continual stages (Fig. 2, Sec. 1)
    The QPA mechanism assumes features retain semantic priors after finetuning; this is shown only qualitatively via K-means on a few images.
  • domain assumption Overlap protocol: the same image can appear in different stages with different labels (Sec. 8)
    CSL (Eq. 6) needs the previous stage's feature map for the current image; disjoint protocol makes it undefined.
  • standard math Mask2Former decoder equations (Eq. 2) as meta-architecture (Sec. 3.2)
    The method builds on Mask2Former's cross-attention, self-attention and FFN decoder; accepted as prior architecture.
  • standard math Bipartite matching from DETR/Mask2Former is used to assign queries to ground truth (Eq. 8)
    VQ storage relies on bipartite matching; taken from prior method [4].
invented entities (2)
  • Virtual Query Bank (B_vq) independent evidence
    purpose: Stores matched decoder query features per class to replay class semantics without storing images.
    Evaluated in Table 5; it is an algorithmic buffer, not a physical entity, but the paper claims storage efficiency and order robustness for it.
  • Trainable per-class prototypes P^t
    purpose: Guide selection of top-K feature points for query pre-alignment.
    Prototypes are internal to the method; they are only tested within the system, so they have no falsifiable handle outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethinking Query-based Transformer for Continual Image Segmentation." pith.science (2026). https://pith.science/paper/3Q7AQEXM

@misc{pith2026250707831,
  author       = {Pith},
  title        = {Pith review of: Rethinking Query-based Transformer for Continual Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3Q7AQEXM}},
  note         = {Machine review of arXiv:2507.07831}
}
read the original abstract

Class-incremental/Continual image segmentation (CIS) aims to train an image segmenter in stages, where the set of available categories differs at each stage. To leverage the built-in objectness of query-based transformers, which mitigates catastrophic forgetting of mask proposals, current methods often decouple mask generation from the continual learning process. This study, however, identifies two key issues with decoupled frameworks: loss of plasticity and heavy reliance on input data order. To address these, we conduct an in-depth investigation of the built-in objectness and find that highly aggregated image features provide a shortcut for queries to generate masks through simple feature alignment. Based on this, we propose SimCIS, a simple yet powerful baseline for CIS. Its core idea is to directly select image features for query assignment, ensuring "perfect alignment" to preserve objectness, while simultaneously allowing queries to select new classes to promote plasticity. To further combat catastrophic forgetting of categories, we introduce cross-stage consistency in selection and an innovative "visual query"-based replay mechanism. Experiments demonstrate that SimCIS consistently outperforms state-of-the-art methods across various segmentation tasks, settings, splits, and input data orders. All models and codes will be made publicly available at https://github.com/SooLab/SimCIS.

Figures

Figures reproduced from arXiv: 2507.07831 by the authors.

Figure 1
Figure 1. Boxplots of PQ metric for our SimCIS and previous SOTA [43] on ADE20K. We train each model on randomly shuf￾fled continual data input orders and report average PQ for base and novel classes. We observe that recent query-based transform￾ers suffer from a loss of plasticity (low average PQ) and heavy reliance on the input data order (high variance). tending beyond classification, continual image segmenta￾tion adapts t… view at source ↗
Figure 2
Figure 2. Clustering results from feature map. Pixel feature pro￾vides sufficient semantic priors (Person) even after finetuning. • The advantage of objectness diminishes and even has a detrimental effect on plasticity as the task sequence shortens. In the shortest two-task setting, they typically achieve performance comparable to or even slightly lower than the baseline. • The built-in objectness is fragile and lacks robustn… view at source ↗
Figure 4
Figure 4. The Overall Architecture of our SimCIS: a lazy Query Pre-Alignment (Sec 4.1) with a Consistent Selection loss (Sec 4.2) to ensure built-in objectness inner and across stages, and Virtual Query (Sec 4.3) to avoid catastrophic forgetting in class prediction. where l denotes the multi-scale layer, D represents the hid￾den dimension, \mathcal {F}_{(l,h,w)} refers to the feature point at po￾sition (h, w) on the l -th lay… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Qualitative comparisons between SimCIS and BalConpas [13] on the ADE20K 100-5 continual panoptic segmentation scenario. Our SimCIS demonstrates significant results, highlighting the effectiveness of our strategies. Step1 Step5 Step10 GT Step3 Step6 Step11 GT [PITH_FUL…
Figure 6
Figure 6. Figure 6: Qualitative examples in continual learning. Reply Num Disk 100-5 (11 tasks) Type Samples Memory base all Image 0 (*20) 0.0MB 35.7 31.8 75 (*20) 3.4MB 38.9 33.4 150 (*20) 6.1MB 38.9 34.0 300 (*20) 11.8MB 38.5 33.7 600 (*20) 21.9MB 39.2 34.3 Virtual Query 0 (*150) 0.0MB …
Figure 7
Figure 7. Figure 7: Qualitative comparisons between SimCIS and BalConpas [13] on the ADE20K 100-5 continual semantic segmentation [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Clustering results comparison between SimCIS and BalConpas. Our SimCIS maintains the semantic priors in the pixel feature [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

95 extracted references · 58 canonical work pages

  1. [2]

    Decomposed knowledge distilla- tion for class-incremental semantic segmentation

    Donghyeon Baek, Youngmin Oh, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Decomposed knowledge distilla- tion for class-incremental semantic segmentation. Advances in Neural Information Processing Systems, 35:10380–10392,

  2. [3]

    Cascade r-cnn: Delv- ing into high quality object detection

    Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delv- ing into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 6154–6162, 2018. 3

    2018

  3. [4]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European confer- ence on computer vision, pages 213–229. Springer, 2020. 5

    2020

  4. [5]

    End-to-end incre- mental learning

    Francisco M Castro, Manuel J Mar ´ın-Jim´enez, Nicol´as Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incre- mental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233–248, 2018. 1, 3

    2018

  5. [6]

    Modeling the background for incremental learning in semantic segmentation

    Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233–9242, 2020. 3, 5, 6, 7, 14

    2020

  6. [7]

    Com- former: Continual learning in semantic and panoptic seg- mentation

    Fabio Cermelli, Matthieu Cord, and Arthur Douillard. Com- former: Continual learning in semantic and panoptic seg- mentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3010– 3020, 2023. 3, 5, 6, 7, 14

    2023

  7. [8]

    Ssul: Semantic segmentation with unknown label for exemplar- based class-incremental learning

    Sungmin Cha, YoungJoon Yoo, Taesup Moon, et al. Ssul: Semantic segmentation with unknown label for exemplar- based class-incremental learning. Advances in neural infor- mation processing systems, 34:10919–10930, 2021. 1, 3, 5, 6, 7

    2021

  8. [9]

    Ssul: Semantic segmentation with unknown label for exemplar- based class-incremental learning

    Sungmin Cha, YoungJoon Yoo, Taesup Moon, et al. Ssul: Semantic segmentation with unknown label for exemplar- based class-incremental learning. Advances in neural infor- mation processing systems, 34:10919–10930, 2021. 16

    2021

Show all 95 references
  1. [10]

    Riemannian walk for incremen- tal learning: Understanding forgetting and intransigence

    Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajan- than, and Philip HS Torr. Riemannian walk for incremen- tal learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532–547, 2018. 1, 3

    2018

  2. [11]

    Efficient lifelong learning with a- gem

    Arslan Chaudhry, Marc’Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a- gem. arXiv preprint arXiv:1812.00420, 2018. 3

    2018 arXiv

  3. [12]

    A survey on graph neural networks and graph transformers in computer vision: A task-oriented perspective

    Chaoqi Chen, Yushuang Wu, Qiyuan Dai, Hong-Yu Zhou, Mutian Xu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. A survey on graph neural networks and graph transformers in computer vision: A task-oriented perspective. IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024. 1

    2024

  4. [13]

    Strike a balance in continual panoptic segmentation, 2024

    Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho Shing Ip, and Sam Kwong. Strike a balance in continual panoptic segmentation, 2024. 3, 5, 6, 7, 8, 15, 18

    2024

  5. [14]

    Semantic image segmentation with deep convolutional nets and fully connected crfs

    Liang-Chieh Chen. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062, 2014. 3

    2014 arXiv

  6. [15]

    Rethinking atrous convolution for semantic image segmentation

    Liang-Chieh Chen. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017

    2017 arXiv

  7. [16]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834–...

    2017

  8. [17]

    Spgnet: Semantic prediction guidance for scene parsing

    Bowen Cheng, Liang-Chieh Chen, Yunchao Wei, Yukun Zhu, Zilong Huang, Jinjun Xiong, Thomas S Huang, Wen- Mei Hwu, and Honghui Shi. Spgnet: Semantic prediction guidance for scene parsing. In Proceedings of the IEEE/CVF international conference on computer vision , pages 5218– 52...

    2019

  9. [18]

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

    Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per- pixel classification is not all you need for semantic segmen- tation. Advances in Neural Information Processing Systems, 34:17864–17875, 2021. 3

    2021

  10. [19]

    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. 1, 3, 6, 7, 15, 16

    2022

  11. [20]

    Curriculum point prompting for weakly-supervised referring image segmentation

    Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13711–13722, 2024. 1

    2024

  12. [21]

    Learning without mem- orizing

    Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without mem- orizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138–5146,

  13. [22]

    Podnet: Pooled outputs dis- tillation for small-tasks incremental learning

    Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs dis- tillation for small-tasks incremental learning. In Computer vision–ECCV 2020: 16th European conference, Glasgow, UK, August 23–28, 2020, proceedings, part XX 16 , pag...

    2020

  14. [23]

    Plop: Learning without forgetting for contin- ual semantic segmentation

    Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for contin- ual semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 4040–4050, 2021. 3, 5, 6, 7

    2021

  15. [24]

    Dytox: Transformers for continual learning with dynamic token expansion

    Arthur Douillard, Alexandre Ram ´e, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285–9295, 2022. 3

    2022

  16. [25]

    Bacs: Background aware continual semantic segmentation

    Mostafa ElAraby, Ali Harakeh, and Liam Paull. Bacs: Background aware continual semantic segmentation. arXiv preprint arXiv:2404.13148, 2024. 3

    2024 arXiv

  17. [26]

    The pascal visual object classes (voc) challenge

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303–338, 2010. 14

    2010

  18. [27]

    Catastrophic forgetting in connectionist networks

    Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128–135, 1999. 3

    1999

  19. [28]

    Multi-evidence filtering and fusion for multi-label classification, object de- tection and semantic segmentation based on weakly super- vised learning

    Weifeng Ge, Sibei Yang, and Yizhou Yu. Multi-evidence filtering and fusion for multi-label classification, object de- tection and semantic segmentation based on weakly super- vised learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, page...

  20. [29]

    Continual segmentation with disentangled objectness learn- ing and class recognition

    Yizheng Gong, Siyue Yu, Xiaoyang Wang, and Jimin Xiao. Continual segmentation with disentangled objectness learn- ing and class recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3848–3857, 2024. 1, 3, 5, 7, 14, 16

    2024

  21. [30]

    Attribution-aware weight transfer: A warm- start initialization for class-incremental semantic segmenta- tion

    Dipam Goswami, Ren ´e Schuster, Joost van de Weijer, and Didier Stricker. Attribution-aware weight transfer: A warm- start initialization for class-incremental semantic segmenta- tion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3...

    2023

  22. [31]

    Simultaneous detection and segmentation

    Bharath Hariharan, Pablo Arbel ´aez, Ross Girshick, and Ji- tendra Malik. Simultaneous detection and segmentation. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13, pages 297–312. Springer, 2014. 3

    2014

  23. [32]

    Clustering algorithms

    JA Hartigan. Clustering algorithms. John Wiley google schola, 2:25–47, 1975. 15

    1975

  24. [33]

    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. 6

    2016

  25. [34]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017. 3

    2017

  26. [35]

    Non-local context encoder: Robust biomedical image segmentation against adversarial attacks

    Xiang He, Sibei Yang, Guanbin Li, Haofeng Li, Huiyou Chang, and Yizhou Yu. Non-local context encoder: Robust biomedical image segmentation against adversarial attacks. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 8417–8424, 2019. 1

    2019

  27. [36]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. stat, 1050:9, 2015. 5

    2015

  28. [37]

    Learning a unified classifier incrementally via rebalancing

    Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831–839,

  29. [38]

    Free-bloom: Zero-shot text-to-video gener- ator with llm director and ldm animator

    Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video gener- ator with llm director and ldm animator. Advances in Neural Information Processing Systems, 36:26135–26158, 2023. 1

    2023

  30. [39]

    Mvtokenflow: High-quality 4d content generation using multiview token flow

    Hanzhuo Huang, Yuan Liu, Ge Zheng, Jiepeng Wang, Zhiyang Dou, and Sibei Yang. Mvtokenflow: High-quality 4d content generation using multiview token flow. arXiv preprint arXiv:2502.11697, 2025. 1

    2025 arXiv

  31. [40]

    Ccnet: Criss-cross attention for semantic segmentation

    Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 603–612, 2019. 3

    2019

  32. [41]

    Oneformer: One transformer to rule universal image segmentation

    Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989–2998, 2023. 16

    2023

  33. [42]

    Vi- sual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning. In European Conference on Computer Vision, pages 709–727. Springer, 2022. 3, 15

    2022

  34. [43]

    Eclipse: Efficient continual learning in panoptic segmen- tation with visual prompt tuning

    Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmen- tation with visual prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3346–3356, 2024. 1, 3, 5, 6, 7, 8, 14, 15, 16

    2024

  35. [44]

    Mask dino: Towards a unified transformer-based framework for object detection and segmentation

    Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pa...

    2023

  36. [45]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelli- gence, 40(12):2935–2947, 2017. 3

    2017

  37. [46]

    Structured attention network for re- ferring image segmentation

    Liang Lin, Pengxiang Yan, Xiaoqian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for re- ferring image segmentation. IEEE Transactions on Multime- dia, 24:1922–1932, 2021. 1

    1922

  38. [47]

    Gradient episodic memory for continual learning

    David Lopez-Paz and Marc’Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neu- ral information processing systems, 30, 2017. 3

    2017

  39. [48]

    Packnet: Adding mul- tiple tasks to a single network by iterative pruning

    Arun Mallya and Svetlana Lazebnik. Packnet: Adding mul- tiple tasks to a single network by iterative pruning. In Pro- ceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7765–7773, 2018. 1, 3

    2018

  40. [49]

    Piggy- back: Adapting a single network to multiple tasks by learn- ing to mask weights

    Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggy- back: Adapting a single network to multiple tasks by learn- ing to mask weights. In Proceedings of the European con- ference on computer vision (ECCV), pages 67–82, 2018. 3

    2018

  41. [50]

    Incremental learn- ing techniques for semantic segmentation

    Umberto Michieli and Pietro Zanuttigh. Incremental learn- ing techniques for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 3

    2019

  42. [51]

    Continual semantic segmentation via repulsion-attraction of sparse and disentan- gled latent representations

    Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentan- gled latent representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 1114–1124, 2021. 3

    2021

  43. [52]

    Learning to remember: A synaptic plasticity driven framework for continual learning

    Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jah- nichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11321–11329, 2019. 3

    2019

  44. [53]

    Class similarity weighted knowl- edge distillation for continual semantic segmentation

    Minh Hieu Phan, Son Lam Phung, Long Tran-Thanh, Abdes- selam Bouzerdoum, et al. Class similarity weighted knowl- edge distillation for continual semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 16866–16875, 2022. 3

    2022

  45. [54]

    icarl: Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE con- ference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 1, 3

    2001

  46. [55]

    Catastrophic forgetting, rehearsal and pseudorehearsal

    Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123–146, 1995. 1, 3

    1995

  47. [56]

    Incremental learning for robust visual tracking

    David A Ross, Jongwoo Lim, Ruei-Sung Lin, and Ming- Hsuan Yang. Incremental learning for robust visual tracking. International journal of computer vision, 77:125–141, 2008. 1

    2008

  48. [57]

    Learning representations by back-propagating er- rors

    David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating er- rors. nature, 323(6088):533–536, 1986. 4

    1986

  49. [59]

    Progressive neural networks

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Raz- van Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 1

    2016 arXiv

  50. [60]

    Incrementer: Transformer for class-incremental semantic segmentation with knowl- edge distillation focusing on old class

    Chao Shang, Hongliang Li, Fanman Meng, Qingbo Wu, Heqian Qiu, and Lanxiao Wang. Incrementer: Transformer for class-incremental semantic segmentation with knowl- edge distillation focusing on old class. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

    2023

  51. [61]

    Edadet: Open-vocabulary object detection using early dense alignment

    Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15724–15734, 2023. 1

    2023

  52. [62]

    Logoprompt: Synthetic text im- ages can be good visual prompts for vision-language models

    Cheng Shi and Sibei Yang. Logoprompt: Synthetic text im- ages can be good visual prompts for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2932–2941, 2023. 1

    2023

  53. [63]

    The devil is in the object bound- ary: Towards annotation-free instance segmentation using foundation models

    Cheng Shi and Sibei Yang. The devil is in the object bound- ary: Towards annotation-free instance segmentation using foundation models. arXiv preprint arXiv:2404.11957, 2024. 1

    2024 arXiv

  54. [64]

    Part2object: Hierarchical unsupervised 3d instance segmentation

    Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. In European Conference on Com- puter Vision, pages 1–18. Springer, 2024. 1

    2024

  55. [65]

    Plain-det: A plain multi-dataset object detector

    Cheng Shi, Yuchen Zhu, and Sibei Yang. Plain-det: A plain multi-dataset object detector. In European Conference on Computer Vision, pages 210–226. Springer, 2024. 1

    2024

  56. [66]

    Continual learning with deep generative replay

    Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 1, 3

    2017

  57. [67]

    Calibrating cnns for life- long learning

    Pravendra Singh, Vinay Kumar Verma, Pratik Mazumder, Lawrence Carin, and Piyush Rai. Calibrating cnns for life- long learning. Advances in Neural Information Processing Systems, 33:15579–15590, 2020. 3

    2020

  58. [68]

    Segmenter: Transformer for semantic segmenta- tion

    Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmenta- tion. In Proceedings of the IEEE/CVF international confer- ence on computer vision, pages 7262–7272, 2021. 3

    2021

  59. [69]

    Con- trastive grouping with transformer for referring image seg- mentation

    Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Con- trastive grouping with transformer for referring image seg- mentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 23570– 23580, 2023. 1

    2023

  60. [70]

    Temporal collection and distribution for referring video object segmentation

    Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466–15476, 2023. 1

    2023

  61. [71]

    Lifelong learning algorithms

    Sebastian Thrun. Lifelong learning algorithms. In Learning to learn, pages 181–209. Springer, 1998. 3

    1998

  62. [72]

    Learning to prompt for continual learning

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 139–149,

  63. [73]

    Memory replay gans: Learning to generate new categories without forgetting

    Chenshen Wu, Luis Herranz, Xialei Liu, Joost Van De Wei- jer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting. Advances in neural information processing systems, 31, 2018. 3

    2018

  64. [74]

    Large scale incre- mental learning

    Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incre- mental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 374–382,

  65. [75]

    Endpoints weight fusion for class incremental semantic segmentation

    Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 7204–7213,

  66. [76]

    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. 3

    2021

  67. [77]

    Early preparation pays off: New classifier pre-tuning for class incremental semantic segmen- tation

    Zhengyuan Xie, Haiquan Lu, Jia-wen Xiao, Enguang Wang, Le Zhang, and Xialei Liu. Early preparation pays off: New classifier pre-tuning for class incremental semantic segmen- tation. arXiv preprint arXiv:2407.14142, 2024. 3

    2024 arXiv

  68. [78]

    Early preparation pays off: New classifier pre-tuning for class incremental semantic segmen- tation

    Zhengyuan Xie, Haiquan Lu, Jia-wen Xiao, Enguang Wang, Le Zhang, and Xialei Liu. Early preparation pays off: New classifier pre-tuning for class incremental semantic segmen- tation. In European Conference on Computer Vision, pages 183–201. Springer, 2025. 16

    2025

  69. [79]

    Der: Dy- namically expandable representation for class incremental learning

    Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dy- namically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3014–3023,

  70. [80]

    Bottom-up shift and reasoning for referring im- age segmentation

    Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring im- age segmentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 11266–11275, 2021. 1

    2021

  71. [81]

    Ocnet: Object context network for scene parsing

    Yuhui Yuan, Lang Huang, Jianyuan Guo, Chao Zhang, Xilin Chen, and Jingdong Wang. Ocnet: Object context network for scene parsing. arXiv preprint arXiv:1809.00916, 2018. 3

    2018 arXiv

  72. [82]

    Representation compensation networks for continual semantic segmentation

    Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7053–7064, 2022. 1, 3

    2022

  73. [83]

    Slca: Slow learner with classifier align- ment for continual learning on a pre-trained model

    Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier align- ment for continual learning on a pre-trained model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19148–19158, 2023. 1, 3

    2023

  74. [84]

    Mining unseen classes via regional object- ness: A simple baseline for incremental segmentation

    Zekang Zhang, Guangyu Gao, Zhiyuan Fang, Jianbo Jiao, and Yunchao Wei. Mining unseen classes via regional object- ness: A simple baseline for incremental segmentation. Ad- vances in neural information processing systems, 35:24340– 24353, 2022. 3

    2022

  75. [85]

    Coinseg: Contrast inter-and intra-class representations for incremental segmentation

    Zekang Zhang, Guangyu Gao, Jianbo Jiao, Chi Harold Liu, and Yunchao Wei. Coinseg: Contrast inter-and intra-class representations for incremental segmentation. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 843–853, 2023. 3

    2023

  76. [86]

    Pyramid scene parsing network

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881–2890, 2017. 3

    2017

  77. [87]

    Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models.Advances in Neu- ral Information Processing Systems, 36:5168–5191, 2023

    Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models.Advances in Neu- ral Information Processing Systems, 36:5168–5191, 2023. 1

    2023

  78. [88]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 633–641,

  79. [89]

    Continual semantic segmentation with automatic memory sample selection

    Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Continual semantic segmentation with automatic memory sample selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3082–3092, 2023. 3 Rethinking Query-based Trans...

    2023

  80. [90]

    Pseudocode for our SimCIS In this section, we present the overall workflow of our method in the pseudo-code Algo. 1. At the beginning, we define some modules, functions, and variables. For the cur- rent stage t and the previous stage t − 1, we define the backbone modules f t b...

  81. [91]

    Following previous works [7, 29, 43], we use ADE20k [88] to train and evaluate our model for both continual panoptic segmentation and continual se- mantic segmentation tasks

    More Dataset and Implementation Details Dataset Information. Following previous works [7, 29, 43], we use ADE20k [88] to train and evaluate our model for both continual panoptic segmentation and continual se- mantic segmentation tasks. The ADE20K dataset con- tains 20, 210 tra...

  82. [92]

    As shown in Tab

    Continual Learning with Random Order Experiment Details. As shown in Tab. 7, we conduct ex- tensive experiments on our model and ECLIPSE [43] under the ten random orders (detailed orders shown in Tab. 9), where nine of them were completely randomly generated using the random m...

  83. [93]

    As shown in the Tab

    More Ablation Study for Stop Gradient As we mention in the main text, we apply stop gradient on selected object query QN after the QPA strategy, to ensure that the information in feature mapF is not disrupted during training, keeping the objectness information stable across di...

  84. [94]

    7, we additionally compare our SimCIS with BalConpas [13] in the 100-5 continual semantic seg- mentation task

    More Visualization Results for CSS As shown in Fig. 7, we additionally compare our SimCIS with BalConpas [13] in the 100-5 continual semantic seg- mentation task. In the first, second, and fourth row from Fig. 7, BalConpas encounters misclassification of the TV and lamps. In t...

  85. [95]

    In the multi-scale feature generated by the pixel decoder, we choose the fea- ture with the highest resolution for clustering

    Built-in Objectness Maintenance Detailed clustering implementation. In the multi-scale feature generated by the pixel decoder, we choose the fea- ture with the highest resolution for clustering. To evaluate the quality of objectness information contained in the fea- tures, we ...

  86. [96]

    However, in our proposed Lazy Query Pre-alignment strategy, the query features have rich information

    The Order of Attention Layers In Mask2Former [19], the authors employ a cross then self- attention mechanism, as they argue that query features to the first self-attention layer are image-independent and do not have signals from the image, thus applying self-attention is unlik...

  87. [97]

    To ensure a fair comparison, we adopt the same Mask2Former [19] as our meta-architecture for im- age segmentation

    Discussion, Limitation and Future Work Discussion of the choice of meta-architecture for image segmentation. To ensure a fair comparison, we adopt the same Mask2Former [19] as our meta-architecture for im- age segmentation. However, recent years have witnessed rapid advancemen...

Pith tools

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