Pith. sign in

REVIEW 4 major objections 6 minor 50 references

Unsupervised Instance Segmentation with Superpixels

T0 review · 4 major / 6 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 This paper claims that a no-annotation pipeline combining MultiCut on self-supervised features with superpixel-guided losses outperforms all prior unsupervised instance segmentation methods, reaching 34.5 APbox on COCO.

desk verdict The method is a sensible CutLER-style pipeline with three genuinely new pieces, but the headline COCO numbers are selected on the same val split used for evaluation, so the SOTA margin is not yet established. read the letter →

arxiv 2509.05352 v1 pith:6F7ZLPIT submitted 2025-09-03 cs.CV cs.AI

Cuong Manh Hoang This is my paper · ORCID
classification cs.CVcs.AI
keywords unsupervisedinstancesegmentationself-supervisedlearningsuperpixelsMultiCutmaskfilteradaptivelosszero-shotobjectdetectionDINOfeatures
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 proposes an unsupervised instance segmentation framework that needs no human annotations and no in-domain training data. It generates coarse object masks by running a MultiCut graph-cut algorithm on DINO self-supervised ViT features, then filters them by an inner-versus-edge affinity score. A novel superpixel-guided mask loss—hard labels from coarse masks plus soft labels propagated over a superpixel graph by color affinity—trains a SOLO segmentation network. A self-training stage with an adaptive loss reweights pixels by mask stability across checkpoints. Reported results on COCO val2017 are 34.5 APbox and 14.6 APmask, roughly 9 and 3 points above the previous best unsupervised method.

What carries the argument

The load-bearing objects are: (i) RAMA, a GPU MultiCut algorithm that partitions self-supervised ViT patch features into an arbitrary number of object proposals; (ii) the mask filter R(M) = mean inner affinity minus mean edge affinity, which ranks and selects the top 60% of masks; (iii) the superpixel-guided mask loss Lsgm, whose hard component assigns superpixel labels from coarse masks and whose soft component propagates probabilities over a minimum spanning tree of superpixels using color affinities; and (iv) the adaptive loss Lad, which reweights per-pixel loss by min-max normalized IoU stability across saved checkpoints, discounting boundary pixels.

What would settle it

Train the full pipeline on COCO with the mask filter inverted—keep the lowest-scoring Q% of masks instead of the highest. If AP does not drop substantially, the filter's affinity criterion is not doing the claimed work of selecting better coarse supervision.

Watch

Extended reading notes

Core claim

The central discovery is that two cheap, unlabeled sources of structure—patch-level self-supervised features and low-level superpixels—can replace human masks in training an instance segmentation model. The paper shows that a MultiCut algorithm on DINO features segments all potential objects without a preset count, and a simple affinity score selects high-quality coarse masks. The superpixel-guided loss converts coarse masks into hard superpixel labels and uses color-affinity propagation over a minimum spanning tree to generate soft labels, letting the network learn boundary-accurate masks from noisy supervision. The adaptive loss then weights each predicted mask by its stability across chec

Load-bearing premise

The whole pipeline assumes the coarse masks produced by MultiCut on DINO features and selected by the affinity filter are accurate enough to supervise the segmentation network; if those masks are systematically wrong, the superpixel losses and the reported gains would not transfer.

Editorial extensions

If this is right

  • A model trained with zero annotations can surpass methods that pretrain on extra unlabeled downstream data by a wide margin, e.g., +4.8 APbox over unMORE on COCO val2017.
  • The superpixel-guided loss transfers across different superpixel algorithms (SNIC, GMMSP, LRW, MCG) and different coarse-mask generators (FreeMask, MaskCut, RAMA), implying the gains are not tied to one implementation.
  • The adaptive loss achieves the effect of multi-round self-training in a single round, reducing training time.
  • Plugging the pipeline into the U2Seg universal segmentation framework improves instance, semantic, and panoptic segmentation metrics over the CutLER-based version, suggesting these coarse masks are a stronger pseudo-label source for other tasks.
  • Because training uses only ImageNet images and no boxes or masks, the approach could scale to larger uncurated image collections.

Reading between the lines

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

  • The mask filter's top-60% selection likely skews training toward large, well-separated instances; a per-size breakdown such as COCO APS would reveal whether superpixel guidance trades small-object recall for boundary precision.
  • The soft-loss graph propagates labels through color affinity alone, so instances with texture or illumination gradients that split superpixels may be systematically under-segmented; adding learned low-level affinities could test this.
  • The framework's dependence on DINO features, with Table 13 showing sensitivity to the pretraining dataset, suggests the method inherits DINO's biases such as part-versus-whole parsing, so the gains are not architecture-agnostic.
  • The checkpoint-stability reweighting scheme is a generic wrapper around pseudo-label training and could be applied to other noisy-label settings, such as open-vocabulary detection or semi-supervised segmentation.
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

4 major / 6 minor

Summary. The paper proposes an unsupervised instance segmentation framework that trains a SOLO network using pseudo-labels derived from DINO ViT features and MCG superpixels. Coarse masks are produced by the RAMA MultiCut algorithm, filtered by an affinity-based mask filter (Eq. 1), and then used together with superpixels and color information in a proposed superpixel-guided mask loss Lsgm (Eq. 9), consisting of a hard loss (Eq. 5) and a soft loss (Eq. 8). A self-training stage with an adaptive loss (Eq. 13) based on checkpoint stability refines the predictions. The authors report state-of-the-art zero-shot results on COCO, COCO20K, PASCAL VOC, UVO, KITTI, and SSDD, plus an extension to unsupervised universal segmentation.

Significance. If the reported numbers are reliable, the contributions are practically significant: the framework would improve unsupervised instance segmentation by a large margin (e.g., 34.5 vs. 25.4 APbox and 14.6 vs. 11.7 APmask on COCO val2017) and introduces a reusable superpixel-guided loss and an adaptive self-training loss. The ablations are coherent and each component gives a visible improvement. However, the central empirical claim is currently undermined by the evaluation protocol: the six hyperparameters in Table 12(a)-(f) are selected on COCO val2017, and the headline results in Table 2 are reported on that same split. With no held-out split, no multiple seeds, and no error bars, the margin over prior methods may be substantially inflated by selection. The paper also omits several training details and contains inconsistent baseline descriptions across tables.

major comments (4)
  1. [§4.1, Tables 12(a)-12(f), Table 2] The headline COCO val2017 results are selected on the same split used for evaluation. Table 12(a)-(f) tunes Q, alpha_1, alpha_2, e, epsilon, and d_hat by reporting APmask on COCO val2017, and Table 2 reports the final APbox/APmask on the same COCO val2017 split. This is a test-set selection loop: the model is chosen to maximize the exact metric and split that is later presented as the benchmark. The paper reports no error bars, no multiple seeds, and no held-out evaluation, so the 9.1 APbox / 2.9 APmask gain over unMORE in Table 2 cannot be separated from selection bias. Please report results on a disjoint validation split (e.g., COCO minival or a train/val split), provide variance across seeds, or describe a statistically sound model-selection procedure.
  2. [§4.2, Tables 2 and 4] The zero-shot protocol is described inconsistently. In the text before Table 4, it says 'All models perform zero-shot unsupervised object detection and instance segmentation without training on downstream datasets,' yet Table 4 includes unMORE, MaskDistill, and FreeSOLO, which Table 2 marks with '*' as methods that 'train on extra unlabeled images from the downstream datasets.' This discrepancy affects the interpretation of the UVO results and must be corrected by clearly separating in-domain pretrained baselines from true zero-shot baselines in every table and in the text.
  3. [§3.1, Eq. (1), and §4.3] The mask filter evaluates each mask using the same DINO features (F) that are used by RAMA to create the mask. The paper does not analyze whether the top-Q% selected masks are biased toward particular object sizes, aspect ratios, or semantic classes, even though the entire training pipeline depends on these pseudo-labels. This is load-bearing for the main claim. Please add diagnostics, such as AP by object size (AP_S/AP_M/AP_L) for the pseudo-label distribution, or compare the filter against a feature-independent criterion, to show that the filter does not systematically discard a biased subset of objects.
  4. [§4.1, Implementation Details] The implementation details are too sparse for reproducibility: the paper does not report the optimizer, learning rate, batch size, number of epochs, scheduler, image resolution, MCG superpixel parameters, RAMA parameters, checkpoint interval for adaptive loss, self-training iterations, or total training time. These details are needed to verify the empirical claims and to allow other researchers to reproduce the method.
minor comments (6)
  1. [§4.2, Table 4 text] Typo: 'FeeSOLO' should be 'FreeSOLO'.
  2. [§3.2, Eq. (3) and Eq. (6)] Typo: 'probablity' should be 'probability'. Also, equation (6) uses max over edge weights with alpha_2; the text does not explain why the maximum, rather than sum, is chosen for the global affinity. A short intuition would help.
  3. [§5, Conclusion] Typo: 'corlors' should be 'colors'.
  4. [Figure 2] The figure contains the stray word 'Taye' and is difficult to parse. Please clean the diagram and ensure all labels are properly rendered.
  5. [§4.3, Table 12] The sentence 'The results show that the proposed method consistently surpasses prior approaches, even when less optimal hyperparameter values are employed' is not supported by the table itself, which only reports APmask for different hyperparameter values and does not compare to prior approaches. Either add the comparison or soften the claim.
  6. [§4.3, Table 13] Table 13 is interesting but only reports a single run per data condition. At least one additional seed would help establish that the observed differences are not noise.

Circularity Check

1 steps flagged · score 6.0 of 10

COCO val2017 hyperparameters are selected on the same split used for the headline comparison; the SOTA margin is therefore partly forced by the selection loop.

  1. fitted input called prediction [Section 4.1 (Implementation Details), Table 12(a)-(f), Table 2]
    "The hyperparameter Q = 60%. For segmentation network, we use SOLO model [34] with ResNet-101 [15] as the backbone. For superpixel-guided mask loss, we set hyperparameters α1 = 100, α2 = 200. For adaptive loss, we set hyperparameters e = 3, ϵ = 0.6, d̂ = 3. ... Table 12: Ablation studies for hyperparameters on COCO val2017 [19]. ... Ours 35.3 17.2 18.8 29.1 14.3 14.8 34.5 16.7 18.4 28.6 13.8 14.6"

    Table 12 reports APmask on COCO val2017 for each candidate value of Q, α1, α2, e, ϵ, and d̂; Section 4.1 then fixes these hyperparameters to the values that maximize that same APmask column (Q=60, α1=100, α2=200, e=3, ϵ=0.6, d̂=3). Table 2 reports the resulting APbox/APmask on the same COCO val2017 split as the headline SOTA. Thus the COCO margin over unMORE is not an unbiased prediction of the method; it is the product of selecting the model on the exact evaluation set. The other benchmarks share the same hyperparameters and provide some independent signal, but they do not validate the COCO-specific claim.

full rationale

The only load-bearing circularity is in the evaluation protocol: all hyperparameters are chosen from Table 12's APmask numbers on COCO val2017, and the same split is then used in Table 2 for the headline SOTA. That is a fitted-input-called-prediction loop for the COCO claim. The rest of the pipeline is not circular in the derivation sense: the MultiCut masks and the mask filter both use DINO features, but the filter is a re-ranking heuristic rather than a quantity defined by the target; the superpixel hard/soft losses are self-referential regularizers; and the adaptive loss is conventional self-training. There is no self-citation chain or imported uniqueness theorem. Transfer results on PASCAL/UVO/KITTI/SSDD are external but share the same COCO-selected hyperparameters, so they mitigate but do not eliminate the selection loop. Score 6 reflects partial circularity: the central COCO margin is statistically forced by the same-split selection.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The framework rests on several empirical domain assumptions imported from prior work: DINO features encode object structure, RAMA MultiCut recovers all instances, MCG superpixels align with object boundaries, and checkpoint consistency is a valid reliability signal. The six free parameters are all tuned on COCO val2017, the same benchmark used for headline results. No new physical or conceptual entities are introduced.

free parameters (6)
  • Q = 60%
    Top-Q fraction of masks kept by the mask filter in Eq. (1), selected by APmask on COCO val2017 in Table 12(a).
  • alpha_1 = 100
    Color similarity bandwidth in Eq. (3) for hard loss, tuned on COCO val2017 in Table 12(b).
  • alpha_2 = 200
    Affinity bandwidth in Eq. (6) for soft loss graph propagation, tuned on COCO val2017 in Table 12(c).
  • e = 3
    Number of checkpoints used in adaptive loss reliability score, Eq. (10), tuned on COCO val2017 in Table 12(d).
  • epsilon = 0.6
    Min-max normalization floor in Eq. (11), tuned on COCO val2017 in Table 12(e).
  • d_hat = 3
    Boundary width threshold in Eq. (12) for pixel weighting, tuned on COCO val2017 in Table 12(f).
assumptions (7)
  • domain assumption DINO self-supervised ViT features provide object-level semantic separation in patch embeddings.
    Used throughout Section 3.1; the whole pipeline depends on these features grouping objects without labels.
  • domain assumption RAMA MultiCut on these features can segment all potential object instances without a predefined object count.
    Invoked in Section 3.1; this is the core replacement for MaskCut.
  • domain assumption The mask quality score in Eq. (1), based on inner versus edge pairwise affinities from the same features, reliably identifies good masks.
    Invoked in Section 3.1 with Q=60%; if the score is biased, the training set is biased.
  • domain assumption MCG superpixels consistently align with object boundaries so that superpixel labels derived from coarse masks are reliable.
    Invoked in Section 3.2; the hard loss labels superpixels only if all pixels agree with the coarse mask.
  • domain assumption Minimum spanning tree max-affinity propagation preserves useful long-range relationships for soft labels.
    Invoked in Section 3.2, Eq. (6); the soft loss assumes the tree-based smoothing improves rather than distorts predictions.
  • domain assumption Checkpoint-consistent masks are more reliable, and weighting by this stability improves self-training.
    Invoked in Section 3.3, Eqs. (10)-(13); if stability correlates with systematic error, the adaptive loss reinforces it.
  • domain assumption Hyperparameters tuned on COCO val2017 transfer to all other evaluation datasets.
    The paper does not tune per dataset; the generalization of the tuned values is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Instance Segmentation with Superpixels." pith.science (2026). https://pith.science/paper/6F7ZLPIT

@misc{pith2026250905352,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Instance Segmentation with Superpixels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6F7ZLPIT}},
  note         = {Machine review of arXiv:2509.05352}
}
read the original abstract

Instance segmentation is essential for numerous computer vision applications, including robotics, human-computer interaction, and autonomous driving. Currently, popular models bring impressive performance in instance segmentation by training with a large number of human annotations, which are costly to collect. For this reason, we present a new framework that efficiently and effectively segments objects without the need for human annotations. Firstly, a MultiCut algorithm is applied to self-supervised features for coarse mask segmentation. Then, a mask filter is employed to obtain high-quality coarse masks. To train the segmentation network, we compute a novel superpixel-guided mask loss, comprising hard loss and soft loss, with high-quality coarse masks and superpixels segmented from low-level image features. Lastly, a self-training process with a new adaptive loss is proposed to improve the quality of predicted masks. We conduct experiments on public datasets in instance segmentation and object detection to demonstrate the effectiveness of the proposed framework. The results show that the proposed framework outperforms previous state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2509.05352 by the authors.

Figure 1
Figure 1. (a) Unlabeled image; (b) Coarse mask derived from self-supervised features; (c) Superpixels derived from low-level image features; (d) Final mask [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed framework. High-level image features are initially extracted by a self-supervised Vision Transformer (ViT). Using these [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. 3.3. Adaptive Loss for Self-Training After training, the predicted masks are considerably better than the original coarse masks [32, 33]. Hence, we propose a new adaptive loss for self-training to improve performance of the segmentation network. Inspired by [37], we use holistic stability to evaluate the reliability of the predicted masks. We store e checkpoints of the segmentation model. Each check￾point is saved a… view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Illustration of the superpixel-guided mask loss ( [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]
Figure 4
Figure 4. Figure 4: Qualitative results on the COCO val2017 [19]. (a) CutLER [32]; (b) unMORE [38]; (c) Ours. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: More qualitative results. (a) COCO20K [26]; (b) Pascal VOC [10]; (c) UVO [31]; (d) KITTI [12]; (e) SSDD [44]. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results on COCO val2017 [19]. (a) Without [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results for unsupervised universal image segmentation on COCO val2017 [19]. (a) U2Seg [22]. (b) Ours. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 47 canonical work pages

  1. [1]

    Rama: A rapid multicut algorithm on gpu, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pp

    Abbas, A., Swoboda, P., 2022. Rama: A rapid multicut algorithm on gpu, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pp. 8193–8202

    work page 2022

  2. [2]

    Superpixels and polygons using sim- ple non-iterative clustering, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Achanta, R., Susstrunk, S., 2017. Superpixels and polygons using sim- ple non-iterative clustering, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4651–4660

    work page 2017

  3. [3]

    Arbel ´aez, P., Pont-Tuset, J., Barron, J., Marques, F., Malik, J.,

    work page

  4. [4]

    Arica, S., Rubin, O., Gershov, S., Laufer, S., 2024. Cuvler: Enhanced unsupervised object discoveries through exhaustive self-supervised trans- formers, in: Proceedings of the IEEE /CVF Conference on Computer Vi- sion and Pattern Recognition, pp. 23105–23114

    work page 2024

  5. [5]

    Superpixel segmentation using gaussian mixture model

    Ban, Z., Liu, J., Cao, L., 2018. Superpixel segmentation using gaussian mixture model. IEEE Transactions on Image Processing 27, 4105–4117

    work page 2018

  6. [6]

    Emerging properties in self-supervised vision trans- formers, in: Proceedings of the IEEE /CVF international conference on computer vision, pp

    Caron, M., Touvron, H., Misra, I., J ´egou, H., Mairal, J., Bojanowski, P., Joulin, A., 2021. Emerging properties in self-supervised vision trans- formers, in: Proceedings of the IEEE /CVF international conference on computer vision, pp. 9650–9660

    work page 2021

  7. [7]

    Masked-attention mask transformer for universal image segmentation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R., 2022. Masked-attention mask transformer for universal image segmentation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 1290–1299

    work page 2022

  8. [8]

    Ima- genet: A large-scale hierarchical image database, in: 2009 IEEE confer- ence on computer vision and pattern recognition, Ieee

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L., 2009. Ima- genet: A large-scale hierarchical image database, in: 2009 IEEE confer- ence on computer vision and pattern recognition, Ieee. pp. 248–255

    work page 2009

Show all 50 references
  1. [9]

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.,

  2. [10]

    Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.,

  3. [11]

    Gao, N., Shan, Y ., Wang, Y ., Zhao, X., Yu, Y ., Yang, M., Huang, K.,

  4. [12]

    Are we ready for autonomous driving? the kitti vision benchmark suite, in: 2012 IEEE conference on computer vision and pattern recognition, IEEE

    Geiger, A., Lenz, P., Urtasun, R., 2012. Are we ready for autonomous driving? the kitti vision benchmark suite, in: 2012 IEEE conference on computer vision and pattern recognition, IEEE. pp. 3354–3361

    2012

  5. [13]

    Unsupervised semantic segmentation by distilling feature correspondences, in: International Confer- ence on Learning Representations, pp

    Hamilton, M., Zhang, Z., Hariharan, B., Snavely, N., Free- man, W.T., 2022. Unsupervised semantic segmentation by distilling feature correspondences, in: International Confer- ence on Learning Representations, pp. 6022–6032. URL: https://openreview.net/forum?id=SaKO6z6Hl0c

    2022

  6. [14]

    Mask r-cnn, in: Proceedings of the IEEE international conference on computer vision, pp

    He, K., Gkioxari, G., Doll ´ar, P., Girshick, R., 2017. Mask r-cnn, in: Proceedings of the IEEE international conference on computer vision, pp. 2961–2969

    2017

  7. [15]

    Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778

    2016

  8. [16]

    Dynamic random walk for superpixel segmentation

    Kang, X., Zhu, L., Ming, A., 2020. Dynamic random walk for superpixel segmentation. IEEE Transactions on Image Processing 29, 3871–3884

    2020

  9. [17]

    Gcbanet: A global context boundary- aware network for sar ship instance segmentation

    Ke, X., Zhang, X., Zhang, T., 2022. Gcbanet: A global context boundary- aware network for sar ship instance segmentation. Remote Sensing 14, 2165

    2022

  10. [18]

    Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y ., 2023. Mask dino: Towards a unified transformer-based framework for object detection and segmentation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 3041–3050

    2023

  11. [19]

    Microsoft coco: Common objects in context, in: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, Springer

    Lin, T.Y ., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll´ar, P., Zitnick, C.L., 2014. Microsoft coco: Common objects in context, in: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, Spri...

    2014

  12. [20]

    Liu, Z., Lin, Y ., Cao, Y ., Hu, H., Wei, Y ., Zhang, Z., Lin, S., Guo, B.,

  13. [21]

    Lu, X., Wang, W., Ma, C., Shen, J., Shao, L., Porikli, F., 2019. See more, know more: Unsupervised video object segmentation with co-attention siamese networks, in: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pp. 3623–3632

    2019

  14. [22]

    Unsu- pervised universal image segmentation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Niu, D., Wang, X., Han, X., Lian, L., Herzig, R., Darrell, T., 2024. Unsu- pervised universal image segmentation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 22744– 22754

    2024

  15. [23]

    Coarse-to- fine video instance segmentation with factorized conditional appearance flows

    Qin, Z., Lu, X., Nie, X., Liu, D., Yin, Y ., Wang, W., 2023. Coarse-to- fine video instance segmentation with factorized conditional appearance flows. IEEE/CAA Journal of Automatica Sinica 10, 1192–1208

    2023

  16. [24]

    Learning hierarchical embedding for video instance segmentation, in: Proceedings of the 29th ACM international conference on multimedia, pp

    Qin, Z., Lu, X., Nie, X., Zhen, X., Yin, Y ., 2021. Learning hierarchical embedding for video instance segmentation, in: Proceedings of the 29th ACM international conference on multimedia, pp. 1884–1892

    2021

  17. [25]

    Cuts3d: Cutting semantics in 3d for 2d unsupervised instance segmen- tation

    Sick, L., Engel, D., Hartwig, S., Hermosilla, P., Ropinski, T., 2025. Cuts3d: Cutting semantics in 3d for 2d unsupervised instance segmen- tation. Proceedings of the IEEE /CVF International Conference on Com- puter Vision

    2025

  18. [26]

    Localizing objects with self-supervised transformers and no labels, in: Proceedings of the British Machine Vision Conference (BMVC), pp

    Sim ´eoni, O., Puy, G., V o, H.V ., Roburin, S., Gidaris, S., Bursuc, A., P´erez, P., Marlet, R., Ponce, J., 2021. Localizing objects with self-supervised transformers and no labels, in: Proceedings of the British Machine Vision Conference (BMVC), pp. 568–575

    2021

  19. [27]

    Learn- able tree filter for structure-preserving feature transform

    Song, L., Li, Y ., Li, Z., Yu, G., Sun, H., Sun, J., Zheng, N., 2019. Learn- able tree filter for structure-preserving feature transform. Advances in neural information processing systems 32

    2019

  20. [28]

    Yfcc100m: The new data in multimedia research

    Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J., 2016. Yfcc100m: The new data in multimedia research. Communications of the ACM 59, 64–73

    2016

  21. [29]

    Discovering object masks with transformers for unsupervised semantic segmentation

    Van Gansbeke, W., Vandenhende, S., Van Gool, L., 2022. Discovering object masks with transformers for unsupervised semantic segmentation. arXiv preprint arXiv:2206.06363

    2022 arXiv

  22. [30]

    Attention is all you need

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I., 2017. Attention is all you need. Advances in neural information processing systems 30

    2017

  23. [31]

    Unidentified video objects: A benchmark for dense, open-world segmentation, in: Proceed- ings of the IEEE /CVF international conference on computer vision, pp

    Wang, W., Feiszli, M., Wang, H., Tran, D., 2021. Unidentified video objects: A benchmark for dense, open-world segmentation, in: Proceed- ings of the IEEE /CVF international conference on computer vision, pp. 10776–10785

    2021

  24. [32]

    Cut and learn for unsupervised object detection and instance segmentation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Wang, X., Girdhar, R., Yu, S.X., Misra, I., 2023a. Cut and learn for unsupervised object detection and instance segmentation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 3124–3134

  25. [33]

    Freesolo: Learning to segment objects without annotations, in: Proceedings of the IEEE /CVF conference on computer vision and pattern recognition, pp

    Wang, X., Yu, Z., De Mello, S., Kautz, J., Anandkumar, A., Shen, C., Alvarez, J.M., 2022a. Freesolo: Learning to segment objects without annotations, in: Proceedings of the IEEE /CVF conference on computer vision and pattern recognition, pp. 14176–14186

  26. [34]

    Solov2: Dynamic 16 and fast instance segmentation

    Wang, X., Zhang, R., Kong, T., Li, L., Shen, C., 2020. Solov2: Dynamic 16 and fast instance segmentation. Advances in Neural information process- ing systems 33, 17721–17732

    2020

  27. [35]

    Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut

    Wang, Y ., Shen, X., Yuan, Y ., Du, Y ., Li, M., Hu, S.X., Crowley, J.L., Vaufreydaz, D., 2023b. Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut. IEEE trans- actions on pattern analysis and machine intelligence

  28. [36]

    Noisy boundaries: Lemon or lemon- ade for semi-supervised instance segmentation?, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Wang, Z., Li, Y ., Wang, S., 2022b. Noisy boundaries: Lemon or lemon- ade for semi-supervised instance segmentation?, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16826–16835

  29. [37]

    St ++: Make self- training work better for semi-supervised semantic segmentation, in: Pro- ceedings of the IEEE /CVF conference on computer vision and pattern recognition, pp

    Yang, L., Zhuo, W., Qi, L., Shi, Y ., Gao, Y ., 2022. St ++: Make self- training work better for semi-supervised semantic segmentation, in: Pro- ceedings of the IEEE /CVF conference on computer vision and pattern recognition, pp. 4268–4277

    2022

  30. [38]

    unmore: Unsupervised multi-object segmentation via center-boundary reasoning

    Yang, Y ., Zhang, Z., Yang, B., 2025. unmore: Unsupervised multi-object segmentation via center-boundary reasoning. Forty-second International Conference on Machine Learning

    2025

  31. [39]

    High-speed ship detection in sar images based on a grid convolutional neural network

    Zhang, T., Zhang, X., 2019. High-speed ship detection in sar images based on a grid convolutional neural network. Remote Sensing 11, 1206

    2019

  32. [40]

    A full-level context squeeze-and-excitation roi extractor for sar ship instance segmentation

    Zhang, T., Zhang, X., 2022a. A full-level context squeeze-and-excitation roi extractor for sar ship instance segmentation. IEEE Geoscience and Remote Sensing Letters 19, 1–5

  33. [41]

    Htc + for sar ship instance segmentation

    Zhang, T., Zhang, X., 2022b. Htc + for sar ship instance segmentation. Remote Sensing 14, 2395

  34. [42]

    A mask attention interaction and scale en- hancement network for sar ship instance segmentation

    Zhang, T., Zhang, X., 2022c. A mask attention interaction and scale en- hancement network for sar ship instance segmentation. IEEE geoscience and remote sensing letters 19, 1–5

  35. [43]

    A polarization fusion network with geo- metric feature embedding for sar ship classification

    Zhang, T., Zhang, X., 2022d. A polarization fusion network with geo- metric feature embedding for sar ship classification. Pattern Recognition 123, 108365

  36. [44]

    Sar ship detection dataset (ssdd): O fficial release and comprehensive data analysis

    Zhang, T., Zhang, X., Li, J., Xu, X., Wang, B., Zhan, X., Xu, Y ., Ke, X., Zeng, T., Su, H., et al., 2021. Sar ship detection dataset (ssdd): O fficial release and comprehensive data analysis. Remote Sensing 13, 3690

    2021

  37. [45]

    Depthwise separable convo- lution neural network for high-speed sar ship detection

    Zhang, T., Zhang, X., Shi, J., Wei, S., 2019. Depthwise separable convo- lution neural network for high-speed sar ship detection. Remote Sensing 11, 2483. 17

    2019

  38. [2010]

    International journal of computer vision 88, 303–338

    The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303–338

  39. [2014]

    Multiscale combinatorial grouping, in: 2014 IEEE Con- ference on Computer Vision and Pattern Recognition, pp. 328–335. doi:10.1109/CVPR.2014.49

    2014 doi

  40. [2019]

    Ssap: Single-shot instance segmentation with a ffinity pyramid, in: Proceedings of the IEEE /CVF international conference on computer vi- sion, pp. 642–651

  41. [2020]

    3493–3402

    An image is worth 16x16 words: Transformers for image recogni- tion at scale, in: International Conference on Learning Representations, pp. 3493–3402. 15

  42. [2021]

    10012–10022

    Swin transformer: Hierarchical vision transformer using shifted windows, in: Proceedings of the IEEE /CVF international conference on computer vision, pp. 10012–10022

Pith tools

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