Pith. sign in

REVIEW 4 major objections 6 minor 54 references

The paper introduces Video-GCD and claims that consistency-aware contrastive learning with memory-guided prototypes discovers novel video categories better than adapted image-based GCD methods.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Video-GCD: a new benchmark and a consistency-aware contrastive learning method for discovering known and novel categories in videos.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection New task (Video-GCD) is genuinely novel and the method is plausible, but the benchmark statistics in Table 1 do not add up, which guts the central empirical claims until fixed. the 4 major comments →

arxiv 2509.06306 v1 pith:Y325NV4T submitted 2025-09-08 cs.CV

Video-based Generalized Category Discovery via Memory-Guided Consistency-Aware Contrastive Learning

Zhang Jing , Pu Nan , Xie Yu Xiang , Guo Yanming , Lu Qianqi , Zou Shiwei , Yan Jie , Chen Yan This is my paper
classification cs.CV
keywords generalized category discoveryvideo category discoverycontrastive learningtemporal consistencymemory-guided representationaction recognitionfine-grained bird classificationopen-world recognition
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper extends Generalized Category Discovery (GCD) to video, where unlabeled clips may belong to known or never-seen categories and static frames are often ambiguous. It builds a benchmark from three action-recognition datasets and two fine-grained bird datasets, and proposes MCCL, a framework that weights contrastive learning by multi-perspective consistency votes and distills class prototypes from both features and logits. The paper claims MCCL beats adapted image-based GCD baselines on four of five datasets, with the largest margin (5.74 points) on fine-grained bird video, and takes this as evidence that temporal cues are necessary for reliable video category discovery. The setting matters because open-world video—from wildlife monitoring to surveillance—cannot be organized by appearance alone.

Core claim

On the paper's own terms, the discovery is that video-based generalized category discovery can be solved by using agreement among spatial, temporal, and spatiotemporal views to supervise contrastive learning, rather than by transplanting image GCD pipelines onto video features. The framework reaches the highest All ACC on UCF101 (68.27), Kinetics-400 (22.95), VB100 (46.17), and IBC127 (38.06), with margins of +4.34, +1.38, +5.74, and +4.47 points over the best baseline; on SSv2 it places second, behind one baseline. The paper reads the overall pattern as evidence that temporal information is essential for discovering novel categories in videos.

What carries the argument

The load-bearing mechanism is two interacting modules. Consistency-Aware Contrastive Learning (CACL) extracts spatial, temporal, and spatiotemporal features, fuses them with residual attention, and gives each instance pair a consistency score by voting over multiple clusterings: three horizontal clusterings on different feature views plus vertical clusterings at different granularities. The scores re-weight an InfoNCE loss so that consensus pairs act as confident positives and disagreed pairs as hard negatives. Memory-Guided Representation Enhancement (MGRE) stores known-class prototypes in a dual-level buffer—features and logits—and adds distillation losses pulling instance features to thei

Load-bearing premise

The method assumes the true total number of categories, including the novel ones, is known in advance and uses that number to set the classifier's output dimension and the K-means cluster count; all headline gains come from this setting.

What would settle it

Run MCCL under the same K-unknown protocol used for the baseline in the paper's Table 6: estimate the category count with the procedure in [55], then train and test MCCL on UCF101, SSv2, VB100, and IBC127. If MCCL's All ACC no longer exceeds the best image-based baseline, the central claim fails in the practical setting the paper itself identifies as realistic.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Video-GCD becomes a distinct benchmark problem: methods must use video backbones and motion-aware objectives rather than applying image GCD losses to single frames.
  • The released five-dataset protocol (UCF101, SSv2, Kinetics-400, VB100, IBC127) with known/unknown splits and All/Old/New ACC metrics gives future video category discovery a common measuring stick.
  • Weighting contrastive pairs by multi-view consistency can be carried into other semi-supervised video tasks where pseudo-labels are noisy.
  • Feature-level and logit-level prototype distillation are complementary: one preserves old-class accuracy, the other improves new-class discovery, and together they raise overall accuracy.
  • Old-class recognition is not fully solved by this method: SSv2 remains below the best baseline, and the paper flags forgetting of known categories as future work.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The K-unknown experiment in the paper reports only the baseline, not MCCL; a direct run of MCCL with estimated category counts would test whether its advantage survives the practical setting where the number of novel categories is unknown.
  • The even-odd known/unknown split may interact with which classes have distinctive motion; random or difficulty-balanced splits would show whether the benchmark's ranking is stable.
  • The dual-level memory idea is not video-specific; applying feature-plus-logit prototype distillation to image GCD, especially fine-grained species datasets, is a natural transfer test.
  • Consistency scores are computed from fixed clusterings; making the voting loop iterative (recluster and re-vote during training) could turn the feedback loop into an online pseudo-label refinement process.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces Video-GCD, a video-domain extension of Generalized Category Discovery, and proposes Memory-guided Consistency-aware Contrastive Learning (MCCL), which combines Consistency-Aware Contrastive Learning (CACL) and Memory-Guided Representation Enhancement (MGRE). The authors construct a new benchmark from five video datasets (UCF101, SSv2, Kinetics-400, VB100, IBC127) with known/unknown category splits, adapt several image-based GCD methods as baselines, and report accuracy gains for MCCL on most datasets. Ablation studies are provided to support the contributions of the individual modules, and a K-unknown experiment is included based on estimated category counts.

Significance. The paper addresses a relevant and under-explored extension of GCD to video, and the proposed method is well motivated: temporal information should help disambiguate actions and fine-grained behaviors that static frames confuse. The authors provide a new benchmark, detailed comparisons against adapted baselines, and component-wise ablations. If the benchmark statistics and evaluation rigor are corrected, the task formulation and method could be a useful contribution to the community. However, the main evidence currently rests on a benchmark whose definitions are internally inconsistent, and the reported gains are not accompanied by error bars or significance tests.

major comments (4)
  1. [Sec. 4.1 / Table 1] The benchmark statistics do not add up. For every dataset, the known-category and unknown-category video counts do not sum to the stated total: UCF101 2,908+9,935=12,843 vs. 13,190; SSv2 5,528+19,086=24,614 vs. 25,253; Kinetics-400 29,516+7,974=37,490 vs. 34,828 (the columns are supposedly disjoint); VB100 1,556+5,539=7,095 vs. 7,280; IBC127 1,840+5,592=7,432 vs. 7,616. The prose also contradicts the table: Sec. 4.1 states UCF101 has 9,935 videos, VB100 has 1,416 clips, and IBC127 has 8,014 clips. Since the exact split composition determines both training and evaluation, the main results in Tables 2-3 are not reproducible and cannot be assigned to a well-defined benchmark. The counts and split protocol must be corrected and reconciled.
  2. [Sec. 4.2 / Tables 2-3] The claim that MCCL 'outperforms other models in all aspects' is not supported by the tables. MCCL's New ACC on Kinetics-400 is 19.51 vs. SelfEx's 20.28, and on IBC127 it is 32.72 vs. SelfEx's 33.50. Several All ACC differences are small (e.g., SSv2: 13.58 vs. 12.54; Kinetics-400: 22.95 vs. 21.57), yet no error bars, multiple seeds, or significance tests are reported. Please provide standard deviations across runs and significance tests, and qualify the superiority claim accordingly.
  3. [Sec. 4.4 / Table 6] The K-unknown experiment is essential because the method explicitly assumes that |C_N| is known (Sec. 3) and uses this value for the classifier output dimension and K-means. However, Table 6 reports results only for the Baseline method, not for MCCL. Thus the paper provides no evidence that MCCL is effective when the number of categories is estimated. Add MCCL results with estimated K for all datasets, or remove the claim that the setting is handled.
  4. [Sec. 3.1 / Eq. (6)] CACL computes consistency scores c_{ij} from clustering assignments of the model's own current features and then uses these weights in the contrastive loss L_HCL. This is a self-training loop that can reinforce initial clustering errors, and the manuscript explicitly describes it as a 'mutually reinforcing feedback loop.' No analysis is given to show that the consistency estimates track ground truth during training. Please add a diagnostic (e.g., consistency-score accuracy vs. true labels over training) or an ablation comparing the current consistency mechanism with oracle or randomly assigned consistency scores, to justify that the feedback is beneficial rather than amplifying noise.
minor comments (6)
  1. [Throughout] Typos: 'firsbee' in Sec. 1, 'debote' in Sec. 4.3, and 'τHTL' in Sec. 4.1 is undefined (likely τ_TL from Eq. 11).
  2. [Table 1] The 'Category Division' entry for VB100, '(0,50),(51,100)', is ambiguous; specify that the first 50 classes are known and the remaining 50 are unknown.
  3. [Table 6] The input names 'Bf eat' and 'Logit' are unclear; define what features are used (e.g., backbone features, logits) and fix the formatting.
  4. [Sec. 4.3] The text reporting CACL improvements as '1.49%, 5.75%, and 2.08%' omits SSv2, where +CACL gives 11.31 vs. baseline 11.36 (a small negative change). Either include this result or state that CACL does not help on SSv2.
  5. [Eq. (7)] The numerator uses τ_H while the denominator uses τ_HI; clarify whether this is intentional or a typo, since the standard InfoNCE formulation uses the same temperature.
  6. [References] References [1] and [55] are the same work (Vaze et al., CVPR 2022). Please merge or distinguish them.

Circularity Check

0 steps flagged

No significant circularity: MCCL's consistency bootstrapping is an algorithmic feedback loop, not a logico-derivational one, and the final evaluation is an independent K-means alignment; self-citations are not load-bearing.

full rationale

The paper's core derivation is the MCCL training objective: CACL computes consistency scores from multi-view clusterings of the model's own features and uses them to weight a contrastive loss, while MGRE distills from feature/logit prototypes. Although the abstract describes a 'mutually reinforcing feedback loop' between consistency modeling and representation learning, this is an iterative self-training/bootstrap mechanism rather than a logical circularity: the reported evaluation is an independent K-means clustering of the final features aligned by Hungarian matching to ground-truth labels (Sec. 3, 'Inference'), so the accuracy numbers are not forced by the consistency weighting by construction. The assumption that |C_N| is known is stated explicitly and is standard in GCD; the K-unknown experiment is acknowledged as a limitation and reports only a baseline, which is a weakness but not a circular step. The self-references in Related Work (e.g., [3,8,11,25]) are not load-bearing; the backbone, losses, and K-estimation are all attributed to external prior work ([21], [38], [55]). The Table 1 dataset-count inconsistencies are a reproducibility/correctness concern, not a circular-derivation concern. No step in the derivation chain reduces to its own input or to a self-citation.

Axiom & Free-Parameter Ledger

9 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or conceptual entities; its load-bearing assumptions are the known-category-count oracle, arbitrary benchmark splits, the sufficiency of TimeSformer features, and the reliability of self-derived clustering consistency weights.

free parameters (9)
  • eta (Eq. 6)
    Trade-off factor between supervised label term and cluster-voting consistency; no value is reported in the paper, chosen by hand.
  • tau_H, tau_HI (Eq. 7) = 1.0
    Contrastive temperatures, following [53]; affect the sharpness of the contrastive distribution.
  • tau_CL (Eq. 9) = 0.05
    Temperature for feature-prototype contrastive loss.
  • tau_TL (Eq. 11) = 0.1
    Sharpening temperature for teacher logits in distillation.
  • tau_SL (Eq. 12)
    Distillation temperature for student logits; no value given.
  • lambda_Sup (Eq. 13) = 0.45
    Supervised loss weight, selected as best on VB100 in Fig. 3(b).
  • lambda_S (Eq. 13) = 0.5
    Weight for logit distillation loss, selected as best on VB100 in Fig. 3(a).
  • Number of voting levels K (Eq. 6)
    Number of clustering levels used in multi-view consistency voting; not specified (fig shows N, N/2, N/4, but total K is unclear).
  • Memory buffer subset size |DM|
    Fraction of labeled data sampled to build the category-level memory buffer; not specified.
axioms (5)
  • domain assumption Total number of categories |C_N| is known in advance and equals the classifier output dimension and K-means cluster count.
    Stated in Section 3 ('we assume that the total number of categories |C_N| is known in advance'); used throughout the main experiments and to compute All/Old/New ACC.
  • ad hoc to paper Even-odd (and 50/50 for VB100) class splits define a fair known/unknown partition for the benchmark.
    The benchmark is constructed by assigning even-indexed classes to known and odd-indexed to unknown; this is an arbitrary and unreleased split whose difficulty is not analyzed.
  • domain assumption TimeSformer's final-block spatiotemporal features are sufficiently discriminative for novel-category clustering.
    The method feeds f_ST to contrastive learning, memory prototypes, and final K-means; if these features are poor for unseen classes, the entire pipeline fails (Section 3.1, Eq. 1).
  • domain assumption Multi-view clustering assignments provide reliable consistency signals for unlabeled pairs.
    CACL weights contrastive pairs by agreement across clusterings of the model's own features; this assumes the clusterings are informative rather than noise (Section 3.1, Eq. 5-6).
  • standard math K-means with Hungarian matching is a valid evaluation of category discovery.
    Evaluation protocol from [1]/[21]; it uses the true number of clusters and optimal label alignment, which can overstate accuracy.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Video-based Generalized Category Discovery via Memory-Guided Consistency-Aware Contrastive Learning." pith.science (2026). https://pith.science/paper/Y325NV4T

@misc{pith2026250906306,
  author       = {Pith},
  title        = {Pith review of: Video-based Generalized Category Discovery via Memory-Guided Consistency-Aware Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y325NV4T}},
  note         = {Machine review of arXiv:2509.06306}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generalized Category Discovery (GCD) is an emerging and challenging open-world problem that has garnered increasing attention in recent years. Most existing GCD methods focus on discovering categories in static images. However, relying solely on static visual content is often insufficient to reliably discover novel categories. To bridge this gap, we extend the GCD problem to the video domain and introduce a new setting, termed Video-GCD. Thus, effectively integrating multi-perspective information across time is crucial for accurate Video-GCD. To tackle this challenge, we propose a novel Memory-guided Consistency-aware Contrastive Learning (MCCL) framework, which explicitly captures temporal-spatial cues and incorporates them into contrastive learning through a consistency-guided voting mechanism. MCCL consists of two core components: Consistency-Aware Contrastive Learning(CACL) and Memory-Guided Representation Enhancement (MGRE). CACL exploits multiperspective temporal features to estimate consistency scores between unlabeled instances, which are then used to weight the contrastive loss accordingly. MGRE introduces a dual-level memory buffer that maintains both feature-level and logit-level representations, providing global context to enhance intra-class compactness and inter-class separability. This in turn refines the consistency estimation in CACL, forming a mutually reinforcing feedback loop between representation learning and consistency modeling. To facilitate a comprehensive evaluation, we construct a new and challenging Video-GCD benchmark, which includes action recognition and bird classification video datasets. Extensive experiments demonstrate that our method significantly outperforms competitive GCD approaches adapted from image-based settings, highlighting the importance of temporal information for discovering novel categories in videos. The code will be publicly available.

Figures

Figures reproduced from arXiv: 2509.06306 by Chen Yan, Guo Yanming, Lu Qianqi, Pu Nan, Xie Yu Xiang, Yan Jie, Zhang Jing, Zou Shiwei.

Figure 1
Figure 1. Figure 1: Schematic diagram of the proposed video-GCD task. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework Overview of our Memory-based Consistency-aware Contrastive Learning [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) and (b) illustrate the impact of hyper-parameters [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 41 canonical work pages

  1. [1]

    Generalized category discovery

    Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. InCVPR, 2022

    work page 2022

  2. [2]

    Learning to discover novel visual categories via deep transfer clustering

    Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. InICCV, 2019

    work page 2019

  3. [3]

    Federated generalized category discovery

    Nan Pu, Wenjing Li, Xingyuan Ji, Yalan Qin, Nicu Sebe, and Zhun Zhong. Federated generalized category discovery. InCVPR, 2024

    work page 2024

  4. [4]

    Happy: A debiased learning framework for continual generalized category discovery.arXiv preprint arXiv:2410.06535, 2024

    Shijie Ma, Fei Zhu, Zhun Zhong, Wenzhuo Liu, Xu-Yao Zhang, and Cheng-Lin Liu. Happy: A debiased learning framework for continual generalized category discovery.arXiv preprint arXiv:2410.06535, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Promptccd: Learning gaussian mixture prompt pool for continual category discovery

    Fernando Julio Cendra, Bingchen Zhao, and Kai Han. Promptccd: Learning gaussian mixture prompt pool for continual category discovery. InEuropean Conference on Computer Vision, pages 188–205. Springer, 2024

    work page 2024

  6. [6]

    Grow and merge: A unified framework for continuous categories discovery.Advances in Neural Information Processing Systems, 35:27455–27468, 2022

    Xinwei Zhang, Jianwen Jiang, Yutong Feng, Zhi-Fan Wu, Xibin Zhao, Hai Wan, Mingqian Tang, Rong Jin, and Yue Gao. Grow and merge: A unified framework for continuous categories discovery.Advances in Neural Information Processing Systems, 35:27455–27468, 2022

    work page 2022

  7. [7]

    Clip-gcd: Simple language guided general- ized category discovery.arXiv preprint arXiv:2305.10420, 2023

    Rabah Ouldnoughi, Chia-Wen Kuo, and Zsolt Kira. Clip-gcd: Simple language guided general- ized category discovery.arXiv preprint arXiv:2305.10420, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    Textual knowledge matters: Cross-modality co-teaching for generalized visual class discovery

    Haiyang Zheng, Nan Pu, Wenjing Li, Nicu Sebe, and Zhun Zhong. Textual knowledge matters: Cross-modality co-teaching for generalized visual class discovery. InECCV, 2024

    work page 2024

  9. [9]

    Multimodal generalized category discovery.arXiv preprint arXiv:2409.11624, 2024

    Yuchang Su, Renping Zhou, Siyu Huang, Xingjian Li, Tianyang Wang, Ziyue Wang, and Min Xu. Multimodal generalized category discovery.arXiv preprint arXiv:2409.11624, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Get: Unlocking the multi-modal potential of clip for generalized category discovery

    Enguang Wang, Zhimao Peng, Zhengyuan Xie, Xialei Liu, and Ming-Ming Cheng. Get: Unlocking the multi-modal potential of clip for generalized category discovery. InCVPR, 2025

    work page 2025

  11. [11]

    Novel class discovery for ultra-fine-grained visual categorization

    Yu Liu, Yaqi Cai, Qi Jia, Binglin Qiu, Weimin Wang, and Nan Pu. Novel class discovery for ultra-fine-grained visual categorization. InCVPR, 2024

    work page 2024

  12. [12]

    Hilo: A learning framework for generalized category discovery robust to domain shifts.arXiv preprint arXiv:2408.04591, 2024

    Hongjun Wang, Sagar Vaze, and Kai Han. Hilo: A learning framework for generalized category discovery robust to domain shifts.arXiv preprint arXiv:2408.04591, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Neighbor- hood contrastive learning for novel class discovery

    Zhun Zhong, Enrico Fini, Subhankar Roy, Zhiming Luo, Elisa Ricci, and Nicu Sebe. Neighbor- hood contrastive learning for novel class discovery. InCVPR, 2021

    work page 2021

  14. [14]

    A unified objective for novel class discovery

    Enrico Fini, Enver Sangineto, Stéphane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. InICCV, 2021

    work page 2021

  15. [15]

    Novel visual category discovery with dual ranking statistics and mutual knowledge distillation

    Bingchen Zhao and Kai Han. Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. InNeurIPS, 2021

    work page 2021

  16. [16]

    Openmix: Reviving known knowledge for discovering novel visual categories in an open world

    Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. InCVPR, 2021. 11

    work page 2021

  17. [17]

    Class-incremental novel class discovery

    Subhankar Roy, Mingxuan Liu, Zhun Zhong, Nicu Sebe, and Elisa Ricci. Class-incremental novel class discovery. InECCV, 2022

    work page 2022

  18. [18]

    Novel class discovery in semantic segmentation

    Yuyang Zhao, Zhun Zhong, Nicu Sebe, and Gim Hee Lee. Novel class discovery in semantic segmentation. InCVPR, 2022

    work page 2022

  19. [20]

    Sinkhorn distances: Lightspeed computation of optimal transport.NeurIPS, 2013

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.NeurIPS, 2013

    work page 2013

  20. [21]

    Parametric classification for generalized category discovery: A baseline study

    Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. InICCV, 2023

    work page 2023

  21. [22]

    No representation rules them all in category discovery

    Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. InNeurIPS, 2024

    work page 2024

  22. [23]

    Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning

    Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. InICLR, 2024

    work page 2024

  23. [24]

    Solving the catastrophic forgetting problem in generalized category discovery

    Xinzi Cao, Xiawu Zheng, Guanhong Wang, Weijiang Yu, Yunhang Shen, Ke Li, Yutong Lu, and Yonghong Tian. Solving the catastrophic forgetting problem in generalized category discovery. InCVPR, 2024

    work page 2024

  24. [25]

    Dynamic conceptional contrastive learning for generalized category discovery

    Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptional contrastive learning for generalized category discovery. InCVPR, 2023

    work page 2023

  25. [26]

    Promptcal: Contrastive affinity learning via auxiliary prompts for general- ized novel category discovery

    Sheng Zhang, Salman Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fa- had Shahbaz Khan. Promptcal: Contrastive affinity learning via auxiliary prompts for general- ized novel category discovery. InCVPR, 2023

    work page 2023

  26. [27]

    Learning semi-supervised gaussian mixture models for generalized category discovery

    Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. InICCV, 2023

    work page 2023

  27. [28]

    Paramet- ric information maximization for generalized category discovery

    Florent Chiaroni, Jose Dolz, Ziko Imtiaz Masud, Amar Mitiche, and Ismail Ben Ayed. Paramet- ric information maximization for generalized category discovery. InICCV, 2023

    work page 2023

  28. [29]

    Contrastive mean-shift learning for generalized category discovery

    Sua Choi, Dahyun Kang, and Minsu Cho. Contrastive mean-shift learning for generalized category discovery. InCVPR, 2024

    work page 2024

  29. [30]

    Videossl: Semi- supervised learning for video classification

    Longlong Jing, Toufiq Parag, Zhe Wu, Yingli Tian, and Hongcheng Wang. Videossl: Semi- supervised learning for video classification. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1110–1119, 2021

    work page 2021

  30. [31]

    Learning from temporal gradient for semi-supervised action recognition

    Junfei Xiao, Longlong Jing, Lin Zhang, Ju He, Qi She, Zongwei Zhou, Alan Yuille, and Yingwei Li. Learning from temporal gradient for semi-supervised action recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3252–3262, 2022

    work page 2022

  31. [32]

    mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412, 2017

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412, 2017

    Pith/arXiv arXiv 2017

  32. [33]

    Cutmix: Regularization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023–6032, 2019

    2019

  33. [34]

    Svformer: Semi-supervised video transformer for action recognition

    Zhen Xing, Qi Dai, Han Hu, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Svformer: Semi-supervised video transformer for action recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18816–18826, 2023

    work page 2023

  34. [35]

    Sefar: Semi-supervised fine-grained action recognition with temporal perturbation and learning stabi- lization

    Yongle Huang, Haodong Chen, Zhenbang Xu, Zihan Jia, Haozhou Sun, and Dian Shao. Sefar: Semi-supervised fine-grained action recognition with temporal perturbation and learning stabi- lization. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 3833–3841, 2025. 12

    work page 2025

  35. [36]

    Timebal- ance: Temporally-invariant and temporally-distinctive video representations for semi-supervised action recognition

    Ishan Rajendrakumar Dave, Mamshad Nayeem Rizve, Chen Chen, and Mubarak Shah. Timebal- ance: Temporally-invariant and temporally-distinctive video representations for semi-supervised action recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2341–2352, 2023

    work page 2023

  36. [37]

    Tclr: Temporal contrastive learning for video representation.Computer Vision and Image Understanding, 219:103406, 2022

    Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. Tclr: Temporal contrastive learning for video representation.Computer Vision and Image Understanding, 219:103406, 2022

    work page 2022

  37. [38]

    Autonovel: Automatically discovering and learning novel visual categories.IEEE TPAMI, 2021

    Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories.IEEE TPAMI, 2021

    work page 2021

  38. [39]

    Is space-time attention all you need for video understanding? InICML, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? InICML, 2021

    2021

  39. [40]

    Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  40. [41]

    Logits deconfusion with clip for few-shot learning

    Shuo Li, Fang Liu, Zehua Hao, Xinyi Wang, Lingling Li, Xu Liu, Puhua Chen, and Wenping Ma. Logits deconfusion with clip for few-shot learning. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 25411–25421, 2025

    work page 2025

  41. [42]

    Visionzip: Longer is better but not necessary in vision language models

    Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 19792–19802, 2025

    work page 2025

  42. [43]

    Fastvlm: Efficient vision encoding for vision language models

    Pavan Kumar Anasosalu Vasu, Fartash Faghri, Chun-Liang Li, Cem Koc, Nate True, Albert Antony, Gokula Santhanam, James Gabriel, Peter Grasch, Oncel Tuzel, et al. Fastvlm: Efficient vision encoding for vision language models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 19769–19780, 2025

    2025

  43. [44]

    A dataset of 101 human action classes from videos in the wild.Center for Research in Computer Vision, 2(11):1–7, 2012

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild.Center for Research in Computer Vision, 2(11):1–7, 2012

    work page 2012

  44. [45]

    something something

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. InProceedings of the IEEE international conference on computer vision, pages 5842– 5850, 2017

    2017

  45. [46]

    Quo vadis, action recognition? a new model and the kinetics dataset

    Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. Inproceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299–6308, 2017

    2017

  46. [47]

    Exploiting temporal information for dcnn-based fine-grained object classification

    ZongYuan Ge, Chris McCool, Conrad Sanderson, Peng Wang, Lingqiao Liu, Ian Reid, and Peter Corke. Exploiting temporal information for dcnn-based fine-grained object classification. In2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA), pages 1–6. IEEE, 2016

    work page 2016

  47. [48]

    Ibc127: Video dataset for fine-grained bird classification

    Tomoaki Saito, Asako Kanezaki, and Tatsuya Harada. Ibc127: Video dataset for fine-grained bird classification. In2016 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2016

    work page 2016

  48. [49]

    Imagenet: A large- scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition. Ieee, 2009

    work page 2009

  49. [50]

    Vivit: A video vision transformer

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu ˇci´c, and Cordelia Schmid. Vivit: A video vision transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 6836–6846, 2021

    2021

  50. [51]

    Is space-time attention all you need for video understanding? InICML, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? InICML, 2021. 13

    work page 2021

  51. [52]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InICCV, 2021

    2021

  52. [53]

    Selex: Self-expertise in fine-grained generalized category discovery

    Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees GM Snoek. Selex: Self-expertise in fine-grained generalized category discovery. InEuropean Conference on Computer Vision, pages 440–458. Springer, 2024

    work page 2024

  53. [54]

    Learn to categorize or categorize to learn? self-coding for generalized category discovery.Advances in Neural Information Processing Systems, 36:72794–72818, 2023

    Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for generalized category discovery.Advances in Neural Information Processing Systems, 36:72794–72818, 2023

    work page 2023

  54. [55]

    Generalized category discovery

    Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7492–7501, 2022. 14

    work page 2022

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.