Pith. sign in

REVIEW 3 major objections 5 minor 54 references

Generalized Class Discovery in Instance Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Making contrastive temperatures, pseudo-label reliability, and attention class-aware advances generalized class discovery in instance segmentation.

desk verdict A solid, well-ablated empirical method for GCD in instance segmentation, with an under-validated headness mechanism and missing reproducibility details. read the letter →

arxiv 2502.08149 v2 pith:STK72AMI submitted 2025-02-12 cs.CV cs.AI

Cuong Manh Hoang , Yeejin Lee , Byeongkeun Kang This is my paper · ORCID
classification cs.CVcs.AI
keywords generalizedcategorydiscoveryinstancesegmentationlong-taileddistributioncontrastivelearningpseudo-labelreliabilitysoftattentionnovelclassopen-world
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

Generalized class discovery (GCD) in instance segmentation asks a model to segment both the classes it was trained on and novel classes it has never seen labeled, using a pool of unlabeled images. This paper argues that the long-tailed distribution of real-world object instances is the main obstacle, and that three mechanisms jointly overcome it: assigning each instance its own contrastive-learning temperature based on a headness score, applying per-class reliability thresholds to pseudo-labels that relax early in training and tighten later, and a soft attention module that focuses the encoder on the target object despite noisy masks. The paper validates the approach on two benchmarks, reporting mAP_all of 12.85 on COCO_half+LVIS and 5.21 on LVIS+Visual Genome, against 6.69 and 4.46 for the previous state of the art. A reader should care because real scene datasets are long-tailed and open-world instance segmentation is a step toward systems that keep learning new categories without exhaustive annotation.

What carries the argument

The central mechanism is the instance-wise temperature assignment (ITA): a headness score $h_i$, computed as the fraction of contrastive similarity mass concentrated in the top-$K$ nearest neighbors of an instance's embedding, is smoothed by momentum and min-max normalized to produce a per-instance temperature $\tau_i$ that replaces the global temperature in the supervised and unsupervised contrastive losses (Eqs. 2–4). The score is a proxy for whether the instance belongs to a frequent head class or a rare tail class, so high temperatures relax instance discrimination for head samples while low temperatures sharpen it for tail samples. The second mechanism, reliability-based dynamic learning (RDL), computes a stability score $s_i$ from the KL divergences between the final model's predictions and those of intermediate checkpoints, ranks samples within each class, and converts the rank into a loss weight $\kappa_t^i$ that starts near 1 for all samples and decays toward 0 for the least stable tail samples as training progresses (Eqs. 10–11). The third mechanism, the soft attention module (SAM), pools feature maps at several scales, computes a pairwise affinity with the full feature map, and multiplies the resulting attention map into the features, trained against masks with reduced weight near object boundaries (Eqs. 5–6). Together, the three components carry the argument that each standard choice—fixed temperature, global reliability threshold, and unguided feature extraction—should be replaced by a class-aware counterpart.

What would settle it

Train the GCD model twice on COCO_half+LVIS, once with the headness score replaced by random temperatures and once with the true class frequency as the score; if the mAP_all gap between ITA and random assignment nearly vanishes, the claim that density-dependent temperatures drive the improvement is falsified, and a low correlation between estimated headness and ground-truth class frequency on the labeled subset would undermine the mechanism even if end-to-end results hold.

Watch

Extended reading notes

Core claim

The paper's central claim is that the standard recipe for GCD in instance segmentation—fixed contrastive temperature, global pseudo-label reliability, and unguided feature extraction—silently penalizes tail classes, and that replacing each piece with a class-aware counterpart produces a large, consistent gain. It claims that the density of a sample's neighborhood in the embedding space, accumulated over epochs, is a usable proxy for headness, and that mapping that score to a per-instance temperature makes head-class samples attract their group while tail-class samples are discriminated individually (Eqs. 2–4). It further claims that pseudo-label reliability, measured by prediction stability across saved checkpoints, should be thresholded per class rather than globally, with thresholds scheduled to admit diverse samples early and only stable samples late (Eqs. 10–11). Finally, it claims that a lightweight attention module trained with boundary-weighted masks suppresses background and neighboring-object interference in the encoder (Eqs. 5–6). With these changes, the method reports mAP_all of 12.85 on COCO_half+LVIS and 5.21 on LVIS+Visual Genome, compared with 6.69 and 4.46 for the previous state of the art; the ablations attribute the gain to all three components.

Load-bearing premise

The load-bearing premise is that the density of an instance's nearest neighbors in the embedding space is a reliable indicator of whether its class is frequent (head) or rare (tail); if the embedding is poorly calibrated, the per-instance temperatures derived from that density could be misassigned and hurt the representation.

Editorial extensions

If this is right

  • Open-world instance segmentation systems can discover thousands of novel classes from unlabeled images while retaining accuracy on known classes, with the largest gains on rare objects.
  • Fixed-temperature contrastive learning leaves a systematic gap on long-tailed data, not only for GCD but for any self-supervised representation learner facing imbalanced classes.
  • Global pseudo-label confidence filtering systematically discards the tail samples that most need labels, so per-class thresholds with a schedule are preferable.
  • The performance is not tied to a single proposal network: swapping in a range of class-agnostic mask generators still keeps the method ahead of the previous state of the art.

Reading between the lines

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

  • Editorial extension: the headness score could be repurposed as a class-prior estimator to set the number of clusters or rank candidate novel classes, since the method already computes a per-instance headness value but only uses it to set temperatures.
  • A neighboring problem this bears on is semi-supervised instance segmentation without novel classes; applying the same class-wise dynamic reliability there would test whether the benefit is specific to GCD or generalizes to any imbalanced pseudo-label training.
  • A testable extension is to vary the pretraining of the backbone while keeping the rest fixed: if per-instance temperatures still help without self-supervised pretraining, the mechanism is about class imbalance rather than a specific initialization.
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 addresses generalized class discovery (GCD) in instance segmentation, where a model must segment both known and novel classes from labeled and unlabeled data under long-tailed instance distributions. The authors propose three components: (1) instance-wise temperature assignment (ITA) for contrastive learning, which estimates a per-instance 'headness' score from neighbor density and assigns higher temperatures to head-class samples; (2) reliability-based dynamic learning (RDL), which uses class-wise stability thresholds and a dynamic curriculum to weight pseudo-labeled instances when training an instance segmentation network; and (3) a soft attention module (SAM) that uses object masks to suppress background features in the embedding backbone. Experiments on COCO_half+LVIS and LVIS+Visual Genome report substantial mAP improvements over prior state-of-the-art RNCDL, and ablations attribute gains to each component. The paper is clearly written and the experimental design is internally consistent, but several load-bearing details are under-validated or ambiguous, and the empirical evidence lacks error bars and a reproducibility artifact.

Significance. If the reported results hold, this would be a meaningful advance for open-world instance segmentation, a harder and more realistic variant of GCD that prior work has only partially addressed. The paper's strengths are its focus on long-tailed instance distributions, the thorough ablation study (Tables 3-7) that isolates the contribution of each proposed module, the evaluation across two complementary settings, and the robustness check across five class-agnostic mask proposal generators. The proposed ITA mechanism, if validated, would also be of independent interest for imbalanced contrastive learning. However, the significance is tempered by the lack of statistical reliability evidence and by the fact that the core ITA mechanism is justified only by a qualitative t-SNE figure rather than by a quantitative test of its assumed relationship to class frequency.

major comments (3)
  1. [Section 3.2, Eq. (2)-(3); Table 3] The explanation for the +ITA gain in Table 3 (Baseline 5.79 vs. +ITA 7.64 mAP_all) depends critically on the claim that the headness score h_i in Eq. (2) is monotonically related to true class frequency, so that high-temperature is assigned to head samples and low-temperature to tail samples. The paper provides only a qualitative t-SNE visualization (Figure 2) and never checks this correspondence on data with known labels. Since h_i is computed from the same embedding that ITA trains, a miscalibrated headness estimate could persist; a tail instance near a large head class could receive an inflated score and thus an incorrectly high temperature. Please provide a quantitative validation, e.g., the Spearman rank correlation between h_i and ground-truth class frequency on the labeled subset, or a controlled experiment that replaces the estimated headness with oracle class-frequency-based temperatures. Without such evidence, the stated mechanism behind the central contrastive contribution is not supported.
  2. [Section 3.3, Eq. (11) and surrounding text] The description of RDL is ambiguous and undermines reproducibility. The text says 'we first compute \bar t_i for each instance i by finding the rank of s_i among the lowest values within its class,' then states 'we assign \gamma to \bar t_i if its proportional rank falls between \gamma/T_{is} and (\gamma+1)/T_{is},' and finally Eq. (11) computes \kappa^t_i from (t - \bar t_i)/T_{is}. It is unclear whether \bar t_i is a rank, an epoch index, a bin label, or a threshold; the notation mixes the model-index \bar t from Eq. (10) with the new per-instance variable. Please rewrite this section with distinct symbols (e.g., rank r_i and bin index b_i) and explicitly state how the rank is converted into the piecewise weight schedule used in Eq. (11).
  3. [Section 4.1 and Tables 1-3] No error bars, seed variance, or multiple runs are reported for any of the quantitative results. The main comparisons in Tables 1 and 2, and especially the ablation gains in Table 3, are the empirical core of the paper; without repeated runs or a statistical test, one cannot assess whether the reported improvements are significant. Given that the per-component gains in Table 3 are 1.85 (ITA) and 3.07 (SAM), and the final gain over RNCDL is 6.16 mAP_all, the absence of variance information is a load-bearing gap in the evidence. Please report mean and standard deviation over at least three seeds, or provide a strong justification for why seed variability is negligible in this setting.
minor comments (5)
  1. [Eq. (10)] The stability score s_i is defined as a sum of inverse KL divergences. When q^{\bar T}_i and q^{\bar t}_i are nearly identical (as expected for stable, easy samples), the inverse becomes very large; if any probability component is exactly zero in one of the distributions, the KL divergence is undefined. Please specify a smoothed KL or add a small epsilon to avoid division by zero and undefined logarithms.
  2. [Tables 1 and 2] The dagger (†) entries for µGCD and NCDLR are described only in the table caption as 'only the GCD model in our method is replaced.' This means they are not full end-to-end baselines but ablated variants within the proposed framework. The main text should state this distinction explicitly, and if full end-to-end results for these methods exist (using their original pipelines), they should be reported for a complete comparison.
  3. [Section 4.1, Implementation Details] The paper says 'The code will be publicly available on GitHub upon publication,' but no anonymous repository or supplementary material is provided for review. Given the intricacy of the SAM and RDL components, a code release or a detailed reproducibility appendix would substantially strengthen the paper.
  4. [Figure 2] The t-SNE figure is used as the key qualitative evidence for ITA's benefit, but the caption is very brief and the axes are not meaningful in a t-SNE plot. It would be more convincing to also show quantitative cluster purity or separation metrics for the embedding with and without ITA.
  5. [Section 3.2, Eq. (3)] The min-max normalization in Eq. (3) uses H_t, the set of headness scores for all samples, but the text does not specify whether H_t is computed over the entire dataset or over the current mini-batch/queue. Since the queue holds only a subset, clarify the exact set used and how this is updated efficiently at every epoch.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central empirical claims rest on held-out evaluation, and the self-referential stability weighting is a training heuristic, not a fitted prediction.

full rationale

The derivation chain is an empirical pipeline, not a closed-form derivation. Eq. (2) defines a headness score from local embedding density and Eq. (3) only min-max normalizes it to a temperature interval; no parameter fitted to the evaluation target is later reported as a prediction. The paper's central comparison (Tables 1 and 2) uses held-out validation images with ground-truth classes matched by the Hungarian algorithm, so the pseudo-labels used in training are never the quantity being measured. The stability weighting in Eqs. (10)-(11) is self-referential (the model's own checkpoints determine pseudo-label reliability), but it is a loss reweighting schedule for fs; the final mAP values depend on external ground truth and therefore the self-reference is not load-bearing. The method borrows standard components (MoCo queue, SimGCD-style parametric classifier, GGN masks, ST++ stability) from prior work without invoking any self-citation chain or imported uniqueness claim. The limitation that the total number of classes is assumed is an explicit input assumption, not a concealed reuse of the output. The unvalidated link between local density and class frequency in Eq. (2) is a robustness/correctness risk, but it does not make the empirical gain circular.

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

The central claim rests on several domain assumptions inherited from prior work: stability-based reliability, the availability of a good class-agnostic mask generator, and the known number of classes. The ITA component introduces a model-specific assumption that embedding density reflects class frequency. The main free parameters are hyperparameters of the proposed modules; the paper reports a single setting without sensitivity analysis or error bars, so the robustness of the method to these choices is not fully established.

free parameters (5)
  • tau_min, tau_max (temperature range for ITA) = tau_min=0.07, tau_max=1
    Eq. (3) uses these bounds for per-instance temperature; Table 4 shows performance varies with the range.
  • lambda (loss balancing weight) = 0.35
    Eq. (9) weights supervised vs unsupervised losses; set following SimGCD (Wen et al. 2023).
  • K (top-K percentage for headness) = 1%
    Eq. (2) uses top-K neighbors to estimate headness; chosen without sensitivity analysis.
  • rho (momentum for headness) = not stated (between 0 and 1)
    Eq. (2) momentum update; value not reported, affecting reproducibility.
  • SAM structural parameters (M, d, s_i, w, d_hat) = M=3, d=D/8, w=0.25*stg, d_hat=1, s1,s2,s3 as defined
    Soft attention module architecture parameters, hand chosen.
assumptions (4)
  • domain assumption The estimation of headness via local density of neighbors in the embedding space (Eq. 2) correctly reflects class frequency, so that assigning high temperature to head samples and low temperature to tail samples improves contrastive learning.
    This is the core premise of ITA; the paper only shows t-SNE for two classes and does not quantitatively verify the density-frequency correspondence.
  • domain assumption The stability of model predictions across training checkpoints (Eq. 10) is a valid proxy for pseudo-label reliability, inherited from ST++ (Yang et al. 2022).
    RDL weights pseudo-labels by stability; if stable predictions are consistently wrong, these samples are kept with high weight.
  • domain assumption The class-agnostic mask generator (GGN) can propose masks for all objects in unlabeled images, including novel tail classes.
    The entire pipeline starts from these masks; if an object is missed, it cannot be discovered or used in training.
  • domain assumption The total number of classes C is known in advance.
    The deep clustering loss (Eq. 7) uses C; the authors acknowledge this assumption in the conclusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalized Class Discovery in Instance Segmentation." pith.science (2026). https://pith.science/paper/STK72AMI

@misc{pith2026250208149,
  author       = {Pith},
  title        = {Pith review of: Generalized Class Discovery in Instance Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/STK72AMI}},
  note         = {Machine review of arXiv:2502.08149}
}
abstract

This work addresses the task of generalized class discovery (GCD) in instance segmentation. The goal is to discover novel classes and obtain a model capable of segmenting instances of both known and novel categories, given labeled and unlabeled data. Since the real world contains numerous objects with long-tailed distributions, the instance distribution for each class is inherently imbalanced. To address the imbalanced distributions, we propose an instance-wise temperature assignment (ITA) method for contrastive learning and class-wise reliability criteria for pseudo-labels. The ITA method relaxes instance discrimination for samples belonging to head classes to enhance GCD. The reliability criteria are to avoid excluding most pseudo-labels for tail classes when training an instance segmentation network using pseudo-labels from GCD. Additionally, we propose dynamically adjusting the criteria to leverage diverse samples in the early stages while relying only on reliable pseudo-labels in the later stages. We also introduce an efficient soft attention module to encode object-specific representations for GCD. Finally, we evaluate our proposed method by conducting experiments on two settings: COCO$_{half}$ + LVIS and LVIS + Visual Genome. The experimental results demonstrate that the proposed method outperforms previous state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2502.08149 by the authors.

Figure 1
Figure 1. Overview of the proposed framework during training. We first train a class-agnostic instance segmentation network [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualization of two semantically close classes. Green: single head class (‘book’); Blue: single tail class (‘booklet’). et al. 2023) showed that a low temperature value tends to uniformly discriminate all instances while a high temper￾ature value leads to group-wise discrimination, as shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the pairwise affinity between the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative results. (a) and (b) are COCO [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 51 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    An, W.; Tian, F.; Zheng, Q.; Ding, W.; Wang, Q.; and Chen, P. 2023. Generalized Category Discovery with Decoupled Prototypical Network. Proceedings of the AAAI Conference on Artificial Intelligence, 37(11): 12527--12535

    work page 2023

  4. [4]

    S.; Maharaj, T.; Fischer, A.; Courville, A.; Bengio, Y.; and Lacoste-Julien, S

    Arpit, D.; Jastrzebski, S.; Ballas, N.; Krueger, D.; Bengio, E.; Kanwal, M. S.; Maharaj, T.; Fischer, A.; Courville, A.; Bengio, Y.; and Lacoste-Julien, S. 2017. A Closer Look at Memorization in Deep Networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, 233--242. PMLR

    work page 2017

  5. [5]

    T.; FENG, Y.; Wu, J.; and Hu, H

    Bai, J.; Liu, Z.; Wang, H.; Chen, R.; Mu, L.; Li, X.; Zhou, J. T.; FENG, Y.; Wu, J.; and Hu, H. 2023. Towards Distribution-Agnostic Generalized Category Discovery. In Advances in Neural Information Processing Systems, volume 36, 58625--58647

    work page 2023

  6. [6]

    Bellver, M.; Salvador, A.; Torrres, J.; and Giro-i Nieto, X. 2019. Budget-aware Semi-Supervised Semantic and Instance Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

    work page 2019

  7. [7]

    Berrada, T.; Couprie, C.; Alahari, K.; and Verbeek, J. 2024. Guided Distillation for Semi-Supervised Instance Segmentation. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 464--472

    work page 2024

  8. [8]

    Cao, K.; Brbic, M.; and Leskovec, J. 2022. Open-World Semi-Supervised Learning. In International Conference on Learning Representations

    work page 2022

Show all 54 references
  1. [9]

    Caron, M.; Touvron, H.; Misra, I.; Jegou, H.; Mairal, J.; Bojanowski, P.; and Joulin, A. 2021. Emerging Properties in Self-Supervised Vision Transformers. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 9630--9640

    2021

  2. [10]

    G.; Kirillov, A.; and Girdhar, R

    Cheng, B.; Misra, I.; Schwing, A. G.; Kirillov, A.; and Girdhar, R. 2022. Masked-attention Mask Transformer for Universal Image Segmentation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 1280--1289

    2022

  3. [11]

    Cuturi, M. 2013. Sinkhorn Distances: Lightspeed Computation of Optimal Transport. In Advances in Neural Information Processing Systems, volume 26

    2013

  4. [12]

    Fini, E.; Sangineto, E.; Lathuilière, S.; Zhong, Z.; Nabi, M.; and Ricci, E. 2021. A Unified Objective for Novel Class Discovery. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 9264--9272

    2021

  5. [13]

    Fomenko, V.; Elezi, I.; Ramanan, D.; Leal-Taix\' e , L.; and Osep, A. 2022. Learning to Discover and Detect Objects. In Advances in Neural Information Processing Systems, volume 35, 8746--8759

    2022

  6. [14]

    Gupta, A.; Dollár, P.; and Girshick, R. 2019. LVIS: A Dataset for Large Vocabulary Instance Segmentation. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 5351--5359

    2019

  7. [15]

    Han, K.; Vedaldi, A.; and Zisserman, A. 2019. Learning to Discover Novel Visual Categories via Deep Transfer Clustering. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 8400--8408

    2019

  8. [16]

    He, K.; Fan, H.; Wu, Y.; Xie, S.; and Girshick, R. 2020. Momentum Contrast for Unsupervised Visual Representation Learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9726--9735

    2020

  9. [17]

    He, K.; Gkioxari, G.; Dollár, P.; and Girshick, R. 2017. Mask R-CNN. In 2017 IEEE International Conference on Computer Vision (ICCV), 2980--2988

    2017

  10. [18]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770--778

    2016

  11. [19]

    Jain, J.; Li, J.; Chiu, M.; Hassani, A.; Orlov, N.; and Shi, H. 2023. OneFormer: One Transformer to Rule Universal Image Segmentation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2989--2998

    2023

  12. [20]

    Kalluri, T.; Wang, W.; Wang, H.; Chandraker, M.; Torresani, L.; and Tran, D. 2024. Open-world Instance Segmentation: Top-down Learning with Bottom-up Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2693--2703

    2024

  13. [21]

    S.; and Kuo, W

    Kim, D.; Lin, T.-Y.; Angelova, A.; Kweon, I. S.; and Kuo, W. 2022. Learning Open-World Object Proposals Without Learning to Classify. IEEE Robotics and Automation Letters, 7(2): 5453--5460

    2022

  14. [22]

    A.; Bernstein, M

    Krishna, R.; Zhu, Y.; Groth, O.; Johnson, J.; Hata, K.; Kravitz, J.; Chen, S.; Kalantidis, Y.; Li, L.-J.; Shamma, D. A.; Bernstein, M. S.; and Fei-Fei, L. 2017. Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations. International Journal of C...

    2017

  15. [23]

    Kuhn, H. W. 1955. The Hungarian Method for the Assignment Problem. Naval Research Logistics Quarterly, 2(1-2): 83--97

    1955

  16. [24]

    Kukleva, A.; Böhle, M.; Schiele, B.; Kuehne, H.; and Rupprecht, C. 2023. Temperature Schedules for Self-Supervised Contrastive Methods on Long-Tail Data. In ICLR

    2023

  17. [25]

    S.; and Anandkumar, A

    Lan, S.; Yu, Z.; Choy, C.; Radhakrishnan, S.; Liu, G.; Zhu, Y.; Davis, L. S.; and Anandkumar, A. 2021. DiscoBox: Weakly Supervised Instance Segmentation and Semantic Correspondence from Box Supervision. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 3386--3396

    2021

  18. [26]

    Li, B.; Yao, Y.; Tan, J.; Zhang, G.; Yu, F.; Lu, J.; and Luo, Y. 2022. Equalized Focal Loss for Dense Long-Tailed Object Detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 6980--6989

    2022

  19. [27]

    Li, Z.; Dai, B.; Simsek, F.; Meinel, C.; and Yang, H. 2023. ImbaGCD: Imbalanced Generalized Category Discovery. arXiv:2401.05353

    2023 arXiv

  20. [28]

    Li, Z.; Meinel, C.; and Yang, H. 2023. Generalized Categories Discovery for Long-tailed Recognition. arXiv:2401.05352

    2023 arXiv

  21. [29]

    Lin, T.-Y.; Goyal, P.; Girshick, R.; He, K.; and Dollár, P. 2017. Focal Loss for Dense Object Detection. In 2017 IEEE International Conference on Computer Vision (ICCV), 2999--3007

    2017

  22. [30]

    MacQueen, J.; et al. 1967. Some Methods for Classification and Analysis of Multivariate Observations. In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, volume 1, 281--297. Oakland, CA, USA

    1967

  23. [31]

    Pu, N.; Zhong, Z.; and Sebe, N. 2023. Dynamic Conceptional Contrastive Learning for Generalized Category Discovery. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 7579--7588

    2023

  24. [32]

    Ren, S.; He, K.; Girshick, R.; and Sun, J. 2015. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In Advances in Neural Information Processing Systems, volume 28

    2015

  25. [33]

    Riz, L.; Saltori, C.; Ricci, E.; and Poiesi, F. 2023. Novel Class Discovery for 3D Point Cloud Semantic Segmentation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9393--9402

    2023

  26. [34]

    Rosvall, M.; and Bergstrom, C. T. 2008. Maps of random walks on complex networks reveal community structure. Proceedings of the National Academy of Sciences, 105(4): 1118--1123

    2008

  27. [35]

    C.; and Fei-Fei, L

    Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; Berg, A. C.; and Fei-Fei, L. 2015. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3): 211--252

    2015

  28. [36]

    Saito, K.; Hu, P.; Darrell, T.; and Saenko, K. 2022. Learning to Detect Every Thing in an Open World. In Computer Vision -- ECCV 2022, 268--284. Cham: Springer Nature Switzerland. ISBN 978-3-031-20053-3

    2022

  29. [37]

    Song, C.; Huang, Y.; Ouyang, W.; and Wang, L. 2018. Mask-Guided Contrastive Attention Model for Person Re-identification. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1179--1188

    2018

  30. [38]

    Vaze, S.; Han, K.; Vedaldi, A.; and Zisserman, A. 2022. Generalized Category Discovery. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 7482--7491

    2022

  31. [39]

    Vaze, S.; Vedaldi, A.; and Zisserman, A. 2023. No Representation Rules Them All in Category Discovery. In Advances in Neural Information Processing Systems, volume 36, 19962--19989

    2023

  32. [40]

    Wang, J.; Yu, X.; and Gao, Y. 2021. Mask Guided Attention For Fine-Grained Patchy Image Classification. In 2021 IEEE International Conference on Image Processing (ICIP), 1044--1048

    2021

  33. [41]

    Wang, W.; Feiszli, M.; Wang, H.; Malik, J.; and Tran, D. 2022. Open-World Instance Segmentation: Exploiting Pseudo Ground Truth From Learned Pairwise Affinity. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 4412--4422

    2022

  34. [42]

    Wang, X.; Zhang, R.; Kong, T.; Li, L.; and Shen, C. 2020. SOLOv2: Dynamic and Fast Instance Segmentation. In Advances in Neural Information Processing Systems, volume 33, 17721--17732

    2020

  35. [43]

    Wang, Z.; Li, Y.; and Wang, S. 2022. Noisy Boundaries: Lemon or Lemonade for Semi-supervised Instance Segmentation? In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 16805--16814

    2022

  36. [44]

    Wen, X.; Zhao, B.; and Qi, X. 2023. Parametric Classification for Generalized Category Discovery: A Baseline Study. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 16544--16554

    2023

  37. [45]

    G.; Limonchik, S.; and Yeung, S

    Weng, Z.; Ogut, M. G.; Limonchik, S.; and Yeung, S. 2021. Unsupervised Discovery of the Long-Tail in Instance Segmentation Using Hierarchical Self-Supervision. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2603--2612

    2021

  38. [46]

    Woo, S.; Park, J.; Lee, J.-Y.; and Kweon, I. S. 2018. CBAM: Convolutional Block Attention Module. In Computer Vision -- ECCV 2018, 3--19. Cham: Springer International Publishing. ISBN 978-3-030-01234-2

    2018

  39. [47]

    Yang, F.; Sun, Q.; Jin, H.; and Zhou, Z. 2020. Superpixel Segmentation With Fully Convolutional Networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 13961--13970

    2020

  40. [48]

    Yang, L.; Li, H.; Wu, Q.; Meng, F.; Qiu, H.; and Xu, L. 2023. Bias-Correction Feature Learner for Semi-Supervised Instance Segmentation. IEEE Transactions on Multimedia, 25: 5852--5863

    2023

  41. [49]

    Yang, L.; Zhuo, W.; Qi, L.; Shi, Y.; and Gao, Y. 2022. ST++: Make Self-training Work Better for Semi-supervised Semantic Segmentation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 4258--4267

    2022

  42. [50]

    Zhang, C.; Xu, R.; and He, X. 2023. Novel Class Discovery for Long-tailed Recognition. Transactions on Machine Learning Research

    2023

  43. [51]

    O.; and Xiang, B

    Zhang, D.; Nan, F.; Wei, X.; Li, S.-W.; Zhu, H.; McKeown, K.; Nallapati, R.; Arnold, A. O.; and Xiang, B. 2021. Supporting Clustering with Contrastive Learning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics...

    2021

  44. [52]

    Zhang, S.; Khan, S.; Shen, Z.; Naseer, M.; Chen, G.; and Khan, F. S. 2023. PromptCAL: Contrastive Affinity Learning via Auxiliary Prompts for Generalized Novel Category Discovery. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 3479--3488

    2023

  45. [53]

    Zhang, Z.; Lan, C.; Zeng, W.; Jin, X.; and Chen, Z. 2020. Relation-Aware Global Attention for Person Re-Identification. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 3183--3192

    2020

  46. [54]

    Zhao, Y.; Zhong, Z.; Sebe, N.; and Lee, G. H. 2022. Novel Class Discovery in Semantic Segmentation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 4330--4339

    2022

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.