Pith. sign in

REVIEW 3 major objections 4 minor 59 references

What Can We Learn from Harry Potter? An Exploratory Study of Visual Representation Learning from Atypical Videos

T0 review · 3 major / 4 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 Training on atypical videos—sci-fi, animation, accidents, surreal scenes—improves open-world video understanding across OOD detection, novel category discovery, and zero-shot action recognition.

desk verdict Useful dataset, plausible diversity story, but the OOD gains on surreal/theatre are likely inflated by train/test overlap. read the letter →

arxiv 2508.21770 v2 pith:BQYNHEMT submitted 2025-08-29 cs.CV

Qiyue Sun , Qiming Huang , Yang Yang , Hongjun Wang , Jianbo Jiao This is my paper · ORCID
classification cs.CV
keywords atypicalvideodatasetopen-worldlearningout-of-distributiondetectionnovelcategorydiscoveryzero-shotactionrecognitionsemanticdiversityauxiliarydatarepresentation
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 asks whether exposing video models to "atypical" footage—sci-fi trailers, animation, unintentional accidents, abnormal events, surreal synthetic scenes, and theatrical performances—during training helps the models cope with open-world inputs. It introduces a new 7,818-clip dataset of such videos and uses them as unlabeled auxiliary data in three open-world tasks: out-of-distribution detection, novel category discovery, and zero-shot action recognition. Across all three tasks, adding atypical data improves over training without auxiliary data and over auxiliary data drawn from conventional datasets like Kinetics-400. The authors' key conclusion is that semantic diversity is the active ingredient: a smaller but more diverse atypical set outperforms a larger, more typical set, and adding more atypical categories steadily improves and stabilizes OOD detection.

What carries the argument

The central object is the newly introduced atypical dataset: 7,818 short clips drawn from six domains—sci-fi, animation, unintentional accidents, abnormal surveillance events, surreal synthetic humans, and theatre—each deliberately far from the everyday human actions in UCF101, HMDB51, Kinetics-400, and Moments-in-Time. The paper uses this unlabeled collection as auxiliary training data in three established pipelines: outlier exposure for OOD detection, AutoNovel's self-supervised-and-clustering recipe for NCD, and ActionCLIP's contrastive video-text training for ZSAR. The load-bearing property is semantic and distributional diversity, measured as information entropy over feature representat

What would settle it

Re-run the OOD experiments with an auxiliary pool that explicitly excludes every clip sourced from the Surreal dataset and all YouTube theatre footage, reusing those only as held-out test sets. If AUROC on [atypical]-surreal and [atypical]-theatre falls back toward the no-auxiliary baseline while the HMDB51 and MiT-v2 gains persist, the claims of open-world generalization to atypical test distributions would need to be revised to claims about exposure to near-duplicates.

Watch

Extended reading notes

Core claim

The paper's central claim is that atypical videos, despite lacking labels and being far from standard action-recognition benchmarks, are a useful resource for open-world video learning. It argues that when these videos are fed into the same straightforward training pipelines used for OOD detection (outlier exposure), novel category discovery (the AutoNovel self-supervised clustering pipeline), and zero-shot action recognition (contrastive video-language alignment), they consistently improve performance on test distributions the model has not seen. The mechanism the paper points to is coverage: the atypical dataset spreads over a wider feature space than typical datasets, and this semantic di

Load-bearing premise

The paper's atypical OOD test sets, [atypical]-surreal and [atypical]-theatre, must be held out of the atypical auxiliary training pool; the paper never states this exclusion explicitly, so the large OOD gains on those two test sets could reflect near-duplicate exposure rather than open-world generalization.

Editorial extensions

If this is right

  • Adding atypical unlabeled videos to OOD detection training lowers FPR95 and raises AUROC on both standard benchmarks (HMDB51, MiT-v2) and atypical test sets, under both 3D-CNN and ViT backbones.
  • For novel category discovery, atypical data combined with the original dataset beats adding Kinetics-400, meaning diversity of auxiliary content matters more than its scale.
  • For zero-shot action recognition, models pre-trained with atypical data generalize better to unseen action classes than those pre-trained on additional typical video.
  • OOD detection performance improves and becomes more stable as more atypical semantic categories are added; no single atypical subset dominates, suggesting complementary coverage.
  • Because the atypical data is unlabeled, the gains are available without annotation cost beyond the initial collection.

Reading between the lines

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

  • This points to a cheap data-side recipe for open-world robustness: collect unusual clips from web trailers, games, or synthetic renders instead of labeling more typical action videos.
  • The paper leaves the style-versus-semantics distinction open: a controlled experiment with stylized-but-real-world actions could determine whether the gains come from unfamiliar appearance, unfamiliar motion, or unfamiliar action semantics.
  • The same logic may transfer to image and audio representations, where atypical auxiliary sets such as surreal or synthetic imagery could supply similar open-set regularization.
  • If semantic diversity is the active ingredient, dataset curation for open-world learning should maximize entropy of features, not just category count; entropy itself could become a selection criterion.
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 / 4 minor

Summary. The paper introduces a new video dataset called "atypical" (7,818 clips from sci-fi, animation, unintentional, abnormal, surreal, and theatre sources) and studies its use as unlabeled auxiliary data in three open-world video tasks: OOD detection, novel category discovery (NCD), and zero-shot action recognition (ZSAR). The authors report that adding atypical data during training improves performance over baselines and over typical auxiliary datasets, and they argue that semantic diversity and content coverage of the auxiliary set are the driving factors. Experiments use ResNet3D-50 and TimeSformer backbones for OOD detection, AutoNovel-style pre-training for NCD, and ActionCLIP for ZSAR.

Significance. If the central claim holds, the paper offers a simple and inexpensive data-side intervention—collecting unusual videos—that reliably improves open-world video understanding. The proposed dataset is also a potentially useful resource. The diversity analysis (Figure 5 and Tables C.4/C.5) is a valuable empirical direction. However, the current evidence is weakened by a likely train/test overlap in the OOD experiments and by the absence of variance estimates, so the headline conclusion is not yet established at the strength claimed. The paper does not ship code or data, and the reported tables appear to be single runs.

major comments (3)
  1. [Section 3, Table 1, Table 2, Table 3, Appendix C.1] Potential OOD leakage: The atypical dataset explicitly includes Surreal (1,024 clips) and Theatre (679 clips) categories (Table 1). In the OOD detection setup, the auxiliary training set is the full atypical set (Table 2), while the OOD test sets include [atypical]-surreal and [atypical]-theatre (Tables 3 and C.1). Appendix C.1 describes removing overlapping action categories only among HMDB51, UCF101, Kinetics-400, and MiT-v2; it never states that surreal and theatre clips are excluded from the auxiliary pool. Since surreal clips come from the Surreal dataset and theatre clips from curated YouTube stage performances, the same videos—or near-duplicates—are very likely present in both training and test. The large AUROC gains on those test sets (75.50→82.03 for surreal; 61.80→74.62 for theatre) could then reflect exposure to near-duplicate test content rather than open-world generalization
  2. [Section 4, Tables 3–5] Missing variance/repeated-seed information. All reported numbers are single values with no error bars or significance tests. Many of the claimed improvements are small (e.g., NCD accuracy 21.99 vs. 20.81 for +atypical vs. +K400 in Table 4; ZSAR Top-1 58.30 vs. 58.23 in Table 5). Without multiple seeds, the central claim that atypical data 'consistently improves' performance is not statistically supported. I request mean ± std over at least three seeds, or equivalent, for the main tables and for the diversity ablations.
  3. [Appendix C.3.2, Table C.2] The statement that HMDB51 NCD results are 'consistent' with UCF101 is contradicted by one of the reported metrics. In Table C.2, +atypical yields NMI 0.1525, which is lower than the baseline NMI of 0.1632 (and also lower than +HMDB51+atypical NMI 0.1922). Accuracy improves (18.18→23.03), but NMI worsens. The paper should either explain why NMI decreases despite accuracy improving, or qualify the claim that atypical data uniformly helps NCD. As written, Section 4.2's conclusion overreaches the HMDB51 evidence.
minor comments (4)
  1. [Appendix C.1] The text says 'we removed 6 overlapping action categories in HMDB51 and UCF101,' but the subsequent list removes 8 categories from HMDB51 and 8 from UCF101. This numerical inconsistency should be corrected.
  2. [Table 2] Table 2 lists Kinetics400, Diving48, and Gaussian Noise under 'Unknown' for OOD detection, but Table 3 reports only HMDB51, MiT-v2, [atypical]-surreal, and [atypical]-theatre. Clarify whether the other unknowns are used as extra test sets or merely as candidate pools. This matters because +K400 is also an auxiliary condition.
  3. [General] The project page is referenced, but no code, dataset download link, or exact clip-level split is provided in the paper. For reproducibility, the authors should release the auxiliary dataset, test-set splits, and training code, or at least give a precise data curation protocol.
  4. [Title / Section 1] The title 'What Can We Learn from Harry Potter?' is evocative but the paper does not explain the connection to the manuscript content; the only Harry Potter reference is a small node in Figure 3. A sentence in the introduction or conclusion would help anchor the title.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; empirical claims rest on controlled comparisons; only minor self-citation, plus a non-circular leakage caveat.

full rationale

This is an empirical study, not a formal derivation. The central claim—that adding atypical auxiliary videos improves OOD detection, NCD, and ZSAR—is supported by controlled experiments against external benchmarks (HMDB51, MiT-v2, UCF101, Kinetics-400) and ablations with non-atypical auxiliary data (Gaussian noise, Diving48). No claimed result is equivalent to its input by construction: the OOD improvements are measured on held-out benchmarks, the NCD/ZSAR gains are evaluated on standard splits, and the diversity analysis keeps total auxiliary sample size fixed while varying semantic categories. The only same-author citation is pace prediction [47], used as the self-supervised pretraining objective within the AutoNovel pipeline for NCD; it is a standard component and is not load-bearing for the atypical-data claim, so it is at most a minor self-citation. The possible overlap between the auxiliary 'atypical' pool and the '[atypical]-surreal' / '[atypical]-theatre' OOD test sets (Table 2: Unknown = [atypical]-surreal, [atypical]-theatre; Auxiliary = atypical) is a data-leakage/validity concern, not circularity: even if confirmed, it would undermine those two OOD rows but not the HMDB51/MiT-v2 OOD results or the NCD/ZSAR findings. Score is 2 for the minor self-citation, not for any reduction of the central claim.

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

Free parameters: none. Experimental hyperparameters are standard choices, not fitted values. Axioms: the central effect relies on transfer from auxiliary atypical data, on the entropy proxy for diversity, and on an undocumented exclusion of surreal/theatre from auxiliary training. Invented entities: none; the atypical dataset is an empirical resource, not a theoretical construction.

assumptions (6)
  • domain assumption Unlabeled auxiliary outliers improve OOD detection via outlier exposure.
    Section 4.1 and Appendix B.1 adopt the outlier exposure objective; the claimed OOD gains depend on this transfer assumption.
  • domain assumption Self-supervised pre-training on auxiliary unlabeled video transfers to NCD and ZSAR.
    Sections 4.2, 4.3 and Appendices B.2 and B.3 use AutoNovel and ActionCLIP pipelines that assume auxiliary data helps downstream generalization.
  • domain assumption Semantic diversity, rather than sample count or visual style, is the active cause of performance gains.
    Section 5.2 fixes sample count while varying categories, but changing categories also changes visual content and domain proximity, so the causal attribution to semantic diversity is an interpretive assumption.
  • domain assumption Information entropy of ResNet3D-50 features is a valid proxy for semantic diversity.
    Section 5.1 and Figure 4 compute entropy after UMAP reduction; the paper does not independently validate this measure as a diversity metric.
  • domain assumption Atypical video categories are disjoint from in-distribution action classes after overlap removal.
    Appendix C.1 removes overlapping categories among HMDB51, UCF101, Kinetics-400, and MiT-v2, but overlap with the atypical subsets is not audited.
  • ad hoc to paper [atypical]-surreal and [atypical]-theatre test sets are excluded from the auxiliary atypical training data.
    Table 2 lists atypical as auxiliary while using these subsets as OOD tests; no explicit exclusion is stated, so the paper tacitly assumes this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of What Can We Learn from Harry Potter? An Exploratory Study of Visual Representation Learning from Atypical Videos." pith.science (2026). https://pith.science/paper/BQYNHEMT

@misc{pith2026250821770,
  author       = {Pith},
  title        = {Pith review of: What Can We Learn from Harry Potter? An Exploratory Study of Visual Representation Learning from Atypical Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BQYNHEMT}},
  note         = {Machine review of arXiv:2508.21770}
}
read the original abstract

Humans usually show exceptional generalisation and discovery ability in the open world, when being shown uncommon new concepts. Whereas most existing studies in the literature focus on common typical data from closed sets, open-world novel discovery is under-explored in videos. In this paper, we are interested in asking: What if atypical unusual videos are exposed in the learning process? To this end, we collect a new video dataset consisting of various types of unusual atypical data (e.g., sci-fi, animation, etc.). To study how such atypical data may benefit open-world learning, we feed them into the model training process for representation learning. Focusing on three key tasks in open-world learning: out-of-distribution (OOD) detection, novel category discovery (NCD), and zero-shot action recognition (ZSAR), we found that even straightforward learning approaches with atypical data consistently improve performance across various settings. Furthermore, we found that increasing the categorical diversity of the atypical samples further boosts OOD detection performance. Additionally, in the NCD task, using a smaller yet more semantically diverse set of atypical samples leads to better performance compared to using a larger but more typical dataset. In the ZSAR setting, the semantic diversity of atypical videos helps the model generalise better to unseen action classes. These observations in our extensive experimental evaluations reveal the benefits of atypical videos for visual representation learning in the open world, together with the newly proposed dataset, encouraging further studies in this direction. The project page is at: https://julysun98.github.io/atypical_dataset.

Figures

Figures reproduced from arXiv: 2508.21770 by the authors.

Figure 1
Figure 1. Illustration of open-world data and tasks. (a) Comparison of open-world [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Representative examples from the proposed [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Composition of the intro￾duced atypical dataset. Existing publicly available datasets primarily focus on common human actions and activities [21, 25, 32, 40]. In contrast, our dataset introduces a broader spectrum of complex and diverse scenarios. As sum￾marised in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Feature distributional en￾tropy comparison between our atyp￾ical dataset and typical datasets. The resulting low-dimensional embeddings are shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Impact of introducing different numbers of semantic categories from the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 51 canonical work pages

  1. [1]

    Ubnor- mal: New benchmark for supervised open-set video anomaly detection

    Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Ubnor- mal: New benchmark for supervised open-set video anomaly detection. InProceed- ings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20143–20153, 2022

    work page 2022

  2. [2]

    Towards open set deep networks

    Abhijit Bendale and Terrance E Boult. Towards open set deep networks. InPro- 20SUN ET AL.: WHA T CAN WE LEARN FROM HARRY POTTER? ceedings of the IEEE conference on computer vision and pattern recognition, pages 1563–1572, 2016

    work page 2016

  3. [3]

    Is space-time attention all you need for video understanding? InProceedings of the International Conference on Machine Learning (ICML), July 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? InProceedings of the International Conference on Machine Learning (ICML), July 2021

    work page 2021

  4. [4]

    Enlarging instance-specific and class-specific information for open-set action recognition

    Jun Cen, Shiwei Zhang, Xiang Wang, Yixuan Pei, Zhiwu Qing, Yingya Zhang, and Qifeng Chen. Enlarging instance-specific and class-specific information for open-set action recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15295–15304, 2023

    work page 2023

  5. [5]

    Elaborative rehearsal for zero-shot action recognition

    Shizhe Chen and Dong Huang. Elaborative rehearsal for zero-shot action recognition. InProceedings of the IEEE/CVF international conference on computer vision, pages 13638–13647, 2021

    2021

  6. [6]

    Wdiscood: Out-of- distribution detection via whitened linear discriminant analysis

    Yiye Chen, Yunzhi Lin, Ruinian Xu, and Patricio A Vela. Wdiscood: Out-of- distribution detection via whitened linear discriminant analysis. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5298–5307, 2023

    work page 2023

  7. [7]

    Haa500: Human-centric atomic action dataset with curated videos

    Jihoon Chung, Cheng-hsin Wuu, Hsuan-ru Yang, Yu-Wing Tai, and Chi-Keung Tang. Haa500: Human-centric atomic action dataset with curated videos. InProceedings of the IEEE/CVF international conference on computer vision, pages 13465–13474, 2021

    work page 2021

  8. [8]

    Towards unknown-aware learning with virtual outlier synthesis

    Xuefeng Du, Zhaoning Wang, Mu Cai, and Sharon Li. Towards unknown-aware learning with virtual outlier synthesis. InInternational Conference on Learn- ing Representations, 2022. URLhttps://openreview.net/forum?id= TW7d65uYu5M

    work page 2022

Show all 59 references
  1. [9]

    Oops! predicting unintentional ac- tion in video

    Dave Epstein, Boyuan Chen, and Carl V ondrick. Oops! predicting unintentional ac- tion in video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 919–929, 2020

    2020

  2. [10]

    Novelcraft: A dataset for novelty detection and discovery in open worlds.arXiv preprint arXiv:2206.11736, 2022

    Patrick Feeney, Sarah Schneider, Panagiotis Lymperopoulos, Li-Ping Liu, Matthias Scheutz, and Michael C Hughes. Novelcraft: A dataset for novelty detection and discovery in open worlds.arXiv preprint arXiv:2206.11736, 2022

    2022 arXiv

  3. [11]

    Unsupervised representation learning by predicting image rotations.arXiv preprint arXiv:1803.07728, 2018

    Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations.arXiv preprint arXiv:1803.07728, 2018

    2018 arXiv

  4. [12]

    Dense open-set recognition with syn- thetic outliers generated by real nvp

    Matej Grci ´c, Petra Bevandi´c, and Siniša Šegvi´c. Dense open-set recognition with syn- thetic outliers generated by real nvp. InInternational Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 133–143, 2021

    2021

  5. [13]

    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. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8401–8409, 2019

    2019

  6. [14]

    Automatically discovering and learning new visual categories with rank- ing statistics

    Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with rank- ing statistics. InInternational Conference on Learning Representations, 2020. SUN ET AL.: WHA T CAN WE LEARN FRO...

    2020

  7. [15]

    Autonovel: Automatically discovering and learning novel visual cate- gories.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10): 6767–6781, 2021

    Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual cate- gories.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10): 6767–6781, 2021

    2021

  8. [16]

    Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? InProceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546–6555, 2018

    Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? InProceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546–6555, 2018

    2018

  9. [17]

    Video owl-vit: Temporally-consistent open-world localization in video

    Georg Heigold, Matthias Minderer, Alexey Gritsenko, Alex Bewley, Daniel Keysers, Mario Luˇci´c, Fisher Yu, and Thomas Kipf. Video owl-vit: Temporally-consistent open-world localization in video. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1...

    2023

  10. [18]

    A baseline for detecting misclassified and out- of-distribution examples in neural networks.International Conference on Learning Representations, 2017

    Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out- of-distribution examples in neural networks.International Conference on Learning Representations, 2017

    2017

  11. [19]

    Deep anomaly detection with outlier exposure.International Conference on Learning Representations, 2019

    Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure.International Conference on Learning Representations, 2019

    2019

  12. [20]

    Generalized odin: De- tecting out-of-distribution image without learning from out-of-distribution data

    Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: De- tecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10951–10960, 2020

    2020

  13. [21]

    The ki- netics human action video dataset.arXiv preprint arXiv:1705.06950, 2017

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The ki- netics human action video dataset.arXiv preprint arXiv:1705.06950, 2017

    2017 arXiv

  14. [22]

    Chal- lenges, evaluation and opportunities for open-world learning.Nature Machine Intelli- gence, 6(6):580–588, 2024

    Mayank Kejriwal, Eric Kildebeck, Robert Steininger, and Abhinav Shrivastava. Chal- lenges, evaluation and opportunities for open-world learning.Nature Machine Intelli- gence, 6(6):580–588, 2024

    2024

  15. [23]

    Opengan: Open-set recognition via open data genera- tion

    Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data genera- tion. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 813–822, 2021

    2021

  16. [24]

    Human action recognition and prediction: A survey.Interna- tional Journal of Computer Vision, 130(5):1366–1401, 2022

    Yu Kong and Yun Fu. Human action recognition and prediction: A survey.Interna- tional Journal of Computer Vision, 130(5):1366–1401, 2022

    2022

  17. [25]

    Hmdb: a large video database for human motion recognition

    Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In2011 Interna- tional conference on computer vision, pages 2556–2563. IEEE, 2011

    2011

  18. [26]

    Resound: Towards action recognition without representation bias

    Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. InProceedings of the European Conference on Computer Vision, pages 513–528, 2018

    2018

  19. [27]

    Resource-rational analysis: Understanding hu- man cognition as the optimal use of limited computational resources.Behavioral and brain sciences, 43:e1, 2020

    Falk Lieder and Thomas L Griffiths. Resource-rational analysis: Understanding hu- man cognition as the optimal use of limited computational resources.Behavioral and brain sciences, 43:e1, 2020. 22SUN ET AL.: WHA T CAN WE LEARN FROM HARRY POTTER?

    2020

  20. [28]

    Abnormal event detection at 150 fps in matlab

    Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 fps in matlab. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2720–2727, 2013

    2013

  21. [29]

    Anomaly de- tection in crowded scenes

    Vijay Mahadevan, Weixin Li, Viral Bhalodia, and Nuno Vasconcelos. Anomaly de- tection in crowded scenes. InProceedings of CVPR, pages 1975–1981, 2010. doi: 10.1109/CVPR.2010.5539872

    1975

  22. [30]

    HowTo100M: Learning a Text-Video Embedding by Watch- ing Hundred Million Narrated Video Clips

    Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watch- ing Hundred Million Narrated Video Clips. InICCV, 2019

    2019

  23. [31]

    Poem: Out-of-distribution detection with pos- terior sampling

    Yifei Ming, Ying Fan, and Yixuan Li. Poem: Out-of-distribution detection with pos- terior sampling. InInternational Conference on Machine Learning, pages 15650– 15665. PMLR, 2022

    2022

  24. [32]

    Mo- ments in time dataset: one million videos for event understanding.IEEE transactions on pattern analysis and machine intelligence, 42(2):502–508, 2019

    Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl V ondrick, et al. Mo- ments in time dataset: one million videos for event understanding.IEEE transactions on pattern analysis and machine int...

    2019

  25. [33]

    Expanding language-image pretrained models for general video recognition

    Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. InEuropean conference on computer vision, pages 1–18. Springer, 2022

    2022

  26. [34]

    Outlier exposure with confidence control for out-of-distribution detec- tion.Neurocomputing, 441:138–150, 2021

    Aristotelis-Angelos Papadopoulos, Mohammad Reza Rajati, Nazim Shaikh, and Ji- amian Wang. Outlier exposure with confidence control for out-of-distribution detec- tion.Neurocomputing, 441:138–150, 2021

    2021

  27. [35]

    A survey on vision-based human action recognition.Image and vision computing, 28(6):976–990, 2010

    Ronald Poppe. A survey on vision-based human action recognition.Image and vision computing, 28(6):976–990, 2010

    2010

  28. [36]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInterna- tional conference on machine learning,...

    2021

  29. [37]

    Fishr: Invariant gradient variances for out-of-distribution generalization

    Alexandre Rame, Corentin Dancette, and Matthieu Cord. Fishr: Invariant gradient variances for out-of-distribution generalization. InInternational Conference on Ma- chine Learning, pages 18347–18377. PMLR, 2022

    2022

  30. [38]

    If deep learning is the answer, what is the question?Nature Reviews Neuroscience, 22(1):55–67, 2021

    Andrew Saxe, Stephanie Nelli, and Christopher Summerfield. If deep learning is the answer, what is the question?Nature Reviews Neuroscience, 22(1):55–67, 2021

    2021

  31. [39]

    Meta- recognition: The theory and practice of recognition score analysis.IEEE transactions on pattern analysis and machine intelligence, 33(8):1689–1695, 2011

    Walter J Scheirer, Anderson Rocha, Ross J Micheals, and Terrance E Boult. Meta- recognition: The theory and practice of recognition score analysis.IEEE transactions on pattern analysis and machine intelligence, 33(8):1689–1695, 2011

    2011

  32. [40]

    Ucf101: A dataset of 101 human actions classes from videos in the wild

    K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. SUN ET AL.: WHA T CAN WE LEARN FROM HARRY POTTER?23

    2012 arXiv

  33. [41]

    Real-world anomaly detection in surveillance videos

    Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479–6488, 2018

    2018

  34. [42]

    Human action recognition from various data modalities: A review.IEEE transactions on pattern analysis and machine intelligence, 45(3):3200–3225, 2022

    Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review.IEEE transactions on pattern analysis and machine intelligence, 45(3):3200–3225, 2022

    2022

  35. [43]

    Black, Ivan Laptev, and Cordelia Schmid

    Gül Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. InCVPR, 2017

    2017

  36. [44]

    Open-set recognition: A good closed-set classifier is all you need? 2021

    Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need? 2021

    2021

  37. [45]

    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

    2022

  38. [46]

    No representation rules them all in category discovery

    Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. InProceedings of the 37th International Conference on Neural Information Processing Systems, pages 19962–19989, 2023

    2023

  39. [47]

    Self-supervised video representation learning by pace prediction

    Jiangliu Wang, Jianbo Jiao, and Yun-Hui Liu. Self-supervised video representation learning by pace prediction. InComputer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVII 16, pages 504–521. Springer, 2020

    2020

  40. [48]

    Action recognition and detection by com- bining motion and appearance features.THUMOS14 Action Recognition Challenge, 1 (2):2, 2014

    Limin Wang, Yu Qiao, Xiaoou Tang, et al. Action recognition and detection by com- bining motion and appearance features.THUMOS14 Action Recognition Challenge, 1 (2):2, 2014

    2014

  41. [49]

    Actionclip: A new paradigm for video action recognition.arXiv preprint arXiv:2109.08472, 2021

    Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition.arXiv preprint arXiv:2109.08472, 2021

    2021 arXiv

  42. [50]

    Openood: Benchmark- ing generalized out-of-distribution detection.Advances in Neural Information Pro- cessing Systems, 35:32598–32611, 2022

    Jingkang Yang, Pengyun Wang, Dejian Zou, Zitang Zhou, Kunyuan Ding, Wenxuan Peng, Haoqi Wang, Guangyao Chen, Bo Li, Yiyou Sun, et al. Openood: Benchmark- ing generalized out-of-distribution detection.Advances in Neural Information Pro- cessing Systems, 35:32598–32611, 2022

    2022

  43. [51]

    Generalized out-of- distribution detection: A survey.International Journal of Computer Vision, pages 1–28, 2024

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of- distribution detection: A survey.International Journal of Computer Vision, pages 1–28, 2024

    2024

  44. [52]

    Un- derstanding deep learning (still) requires rethinking generalization.Communications of the ACM, 64(3):107–115, 2021

    Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Un- derstanding deep learning (still) requires rethinking generalization.Communications of the ACM, 64(3):107–115, 2021

    2021

  45. [53]

    Mixture outlier exposure: Towards out-of-distribution detection in fine-grained en- vironments

    Jingyang Zhang, Nathan Inkawhich, Randolph Linderman, Yiran Chen, and Hai Li. Mixture outlier exposure: Towards out-of-distribution detection in fine-grained en- vironments. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5531–5540, 20...

    2023

  46. [54]

    Open- mix: 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. Open- mix: Reviving known knowledge for discovering novel visual categories in an open world. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9462–9470, 2021

    2021

  47. [55]

    Learning placeholders for open-set recognition

    Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Learning placeholders for open-set recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401–4410, June 2021

    2021

  48. [56]

    Diversified outlier exposure for out-of-distribution detection via in- formative extrapolation.Advances in Neural Information Processing Systems, 36: 22702–22734, 2023

    Jianing Zhu, Yu Geng, Jiangchao Yao, Tongliang Liu, Gang Niu, Masashi Sugiyama, and Bo Han. Diversified outlier exposure for out-of-distribution detection via in- formative extrapolation.Advances in Neural Information Processing Systems, 36: 22702–22734, 2023

    2023

  49. [57]

    Unleashing mask: Explore the intrinsic out-of-distribution detection capa- bility

    Jianing Zhu, Hengzhuang Li, Jiangchao Yao, Tongliang Liu, Jianliang Xu, and Bo Han. Unleashing mask: Explore the intrinsic out-of-distribution detection capa- bility. InInternational Conference on Machine Learning, 2023

    2023

  50. [58]

    Towards universal representation for unseen action recognition

    Yi Zhu, Yang Long, Yu Guan, Shawn Newsam, and Ling Shao. Towards universal representation for unseen action recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 9436–9445, 2018

    2018

  51. [59]

    Towards open set video anomaly detec- tion

    Yuansheng Zhu, Wentao Bao, and Qi Yu. Towards open set video anomaly detec- tion. InProceedings of the European Conference on Computer Vision, pages 395–

Pith tools

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