Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Simplifying Traffic Anomaly Detection with Video Foundation Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a plain encoder-only Video Vision Transformer with self-supervised masked-video pre-training outperforms specialized traffic-anomaly detection models while being substantially more efficient.

desk verdict A solid, replication-friendly empirical study showing that a plain VideoMAE encoder-only model beats specialized TAD architectures, with the caveat that the SOTA and pretraining-objective comparisons are not fully controlled. read the letter →

arxiv 2507.09338 v2 pith:D6JKZLVB submitted 2025-07-12 cs.CV

Svetlana Orlova , Tommie Kerssies , Brunó B. Englert , Gijs Dubbelman This is my paper
classification cs.CV
keywords trafficanomalydetectionvideofoundationmodelsmaskedmodelingdomain-adaptivepre-trainingencoder-onlyarchitectureVisionTransformerDoTADADA-2000
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

This paper sets out to test whether Traffic Anomaly Detection (TAD) needs the complex multi-stage and multi-representation architectures that currently dominate the field. It argues that a single plain Video Vision Transformer, used as an encoder with only a linear classifier on top, can match or beat specialized detectors once it has been pre-trained well. The load-bearing evidence is a set of comparisons in which self-supervised Masked Video Modeling outperforms weakly- and fully-supervised pre-training for TAD, and an extra self-supervised adaptation step on ordinary unlabeled driving video improves results further. If these claims hold, most architectural complexity in current TAD systems is unnecessary on the existing benchmarks, and pre-training strategy becomes the main design lever.

What carries the argument

The central object is the plain Video Vision Transformer (Video ViT), used in an encoder-only design: the video is cut into spatio-temporal tubelets, processed by Transformer blocks, pooled into one feature vector, and followed by a single linear classification layer. The mechanism that carries the argument is Masked Video Modeling (MVM) pre-training, in which the model reconstructs masked spatio-temporal regions, together with an intermediate Domain-Adaptive Pre-Training (DAPT) stage that continues the same reconstruction objective on unlabeled driving videos before fine-tuning. This combination injects motion-sensitive, patch-dense representations into a model with no task-specific inductive bias.

What would settle it

Train three Video ViTs with the same architecture, data, and compute budget, differing only in objective (masked reconstruction, video-text alignment, and class-label classification), then fine-tune all three on DoTA and DADA-2000; if the masked-reconstruction model does not match or beat the others, the paper's central claim about MVM being the strongest signal for TAD is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that pre-training quality, not architecture, decides TAD performance. An encoder-only Video ViT with a single linear head, equipped with Masked Video Modeling pre-training and Domain-Adaptive Pre-Training on unlabeled driving videos, sets a new state of the art on both DoTA and DADA-2000: 88.4 and 88.5 AUCROC, respectively, against 86.2 and 82.7 for the best prior specialized method, while running at 34 frames per second versus 18-33 for the fusion models. The paper further finds that general video-classification accuracy is not predictive of TAD performance, that weak supervision from language and full supervision from class labels generalize poorly to anomalous motion, and that adaptation on anomaly-free driving data is enough to improve detection.

Load-bearing premise

The load-bearing premise is that the ranking of pre-training strategies in Table 3 reflects the training objective itself; the compared checkpoints also differ in pretraining data scale, distillation stages, and training budget, and the paper does not train weakly- or fully-supervised models on the same data and budget as VideoMAE.

Editorial extensions

If this is right

  • Specialized TAD components, such as temporal modules and multi-representation fusion, can be dropped without loss on current DoTA and DADA-2000 benchmarks.
  • The choice of pre-training objective, not benchmark accuracy on Kinetics-400 or Something-SomethingV2, is the reliable predictor of TAD transfer.
  • Anomaly labels are not needed for domain adaptation; ordinary unlabeled driving footage is sufficient to gain most of the benefit.
  • Efficiency improves sharply: the best encoder-only model runs at 34 FPS with a parameter count and GPU memory footprint below or comparable to the specialized baselines.
  • Smaller Video ViTs gain the most from domain-adaptive pre-training, offering an efficient scaling path for real-time deployment.

Reading between the lines

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

  • Beyond the paper, this recipe (MVM pre-training plus DAPT) is a natural baseline for other ego-centric driving tasks that hinge on motion irregularity, such as collision prediction, near-miss detection, and driver attention estimation.
  • Beyond the paper, the absence of correlation between general video benchmarks and TAD suggests that any large-scale evaluation of video foundation models for autonomous driving should include a TAD-style probe alongside standard action recognition.
  • Beyond the paper, a controlled ablation that trains weakly- and fully-supervised models on exactly the same data and compute budget as the MVM model would isolate whether the observed ranking comes from the objective or from checkpoint-side confounds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a simple encoder-only Traffic Anomaly Detection (TAD) pipeline consisting of a plain Video Vision Transformer (Video ViT) followed by a single linear classification head, and studies how pre-training affects performance. On DoTA and DADA-2000, the authors report that DAPT-VideoMAE-L achieves in-domain AUCROC of 88.4% and 88.5%, respectively, outperforming prior specialized TAD methods while claiming high efficiency. The paper further compares off-the-shelf Video ViFMs with different pre-training objectives and concludes that self-supervised Masked Video Modeling (MVM) is the most effective pre-training strategy for TAD, and that domain-adaptive pre-training (DAPT) on unlabeled driving videos improves performance, especially for smaller models, without requiring anomalous examples. The authors release code, domain-adapted encoders, and fine-tuned models.

Significance. If the claims hold, the paper provides a useful and practical result: complex two-stage or multi-representation fusion architectures may be unnecessary on current TAD benchmarks, and pre-training choice matters more than architectural complexity. The paper also offers a label-efficient recipe (DAPT on unlabeled driving data) and releases artifacts that should facilitate reproducibility. Strengths include the controlled comparison among MVM variants in Appendix B, the DAPT ablation in Table 4 that separates domain relevance from additional pre-training, and the explicit reporting of efficiency metrics. However, the two main comparative claims—state-of-the-art performance and the superiority of MVM as an objective—rest on comparisons that are not fully controlled, as detailed below.

major comments (4)
  1. [Sec. 3.3 and Table 3] The comparison in Table 3 does not isolate the pre-training objective. The models differ not only in objective (FSL, WSL, SSL, or combinations) but also in pre-training data scale, distillation stages, input tubelet design, and number of training epochs. For example, VideoMAE2 is listed as SSL+FSL but actually involves dual masking, an additional classification stage, and logit distillation from a larger model, plus about 6x larger pre-training data. Therefore, the top ranking of VideoMAE2 and the general conclusion that MVM is the strongest signal for TAD are confounded. To support contribution (ii), the authors should either train weakly- and fully-supervised models with the same architecture, data, and budget as VideoMAE, or explicitly document and control for these confounds and soften the causal claim.
  2. [Table 2 versus Appendix C, Table 8] The ISCRTAD number reported for DADA-2000 in Table 2 (82.7, marked with a dagger) is inconsistent with the authors' own Appendix C. In Table 8, ISCRTAD's DoTA-to-DADA-2000 generalization performance is broken down as 82.7 for the ego-involved subset, 66.9 for the non-ego subset, and 74.2 overall. The rows for TTHF and PromptTAD in the same Table 8 use the overall (Both) values, matching Table 2 (71.7 and 74.6). Using the ego-only value for ISCRTAD while reporting full-set numbers for the proposed models is an apples-to-oranges comparison and changes the reported margin. Table 2 should either use the overall value 74.2 or explicitly label the subset; the state-of-the-art comparison as currently presented is not reliable.
  3. [Sec. 4.2, Table 2] The claim that the encoder-only Video ViT models are 'significantly more efficient' than top-performing specialized methods is not supported for the top-scoring Large model. Table 2 shows DAPT-VideoMAE-L at 34 FPS versus ISCRTAD at 33 FPS, with ISCRTAD's FPS being an optimistic estimate from public components, and DAPT-VideoMAE-L uses more GPU memory (1.80 GB versus 1.51 GB). The efficiency advantage is clear for the Small and Base variants, but the paper's abstract and conclusion state the efficiency advantage without this qualification. The comparison also mixes measured FPS on an A100 MIG half-GPU with estimates from prior work, so the efficiency numbers should be re-measured under a common protocol or presented with clearer caveats.
  4. [Sec. 3.2 and Table 2] All comparisons to specialized methods rely on numbers taken from prior papers (marked with a dagger) rather than re-running those methods under the authors' evaluation protocol. This is a concern because the protocols differ in input temporal sampling (e.g., 16 frames at 10 FPS for the proposed models vs. frame-by-frame processing at 640x480 for MOVAD), checkpoint selection (highest validation AUCROC), and the use of manually refined annotations. The margin over some prior methods is small enough that protocol differences could affect the ranking. The authors should either re-run representative prior methods under their protocol, or present a sensitivity analysis, and should temper the 'new state of the art' conclusion accordingly.
minor comments (5)
  1. [Sec. 4.4] The sentence describing the DAPT setup is garbled: 'perform several epochs of additional pre-training with the VideoMAE objective on DoTAref, AUCMCC D2K, AUCMCC...' The authors should clarify exactly which data are used for DAPT (Kinetics-700, BDD100K, CAP-DATA, or possibly in-domain unlabeled DoTA) and whether DAPT ever sees labels or the test split.
  2. [Table 3] The type labels in Table 3 are too coarse: VideoMAE2 is a multi-stage pipeline (SSL, then FSL, then logit distillation), SMILE uses masked feature distillation from a vision-language teacher, and InternVideo2 combines several objectives. A single 'Type' column does not capture these differences; the table should list the stages or note them explicitly in the caption.
  3. [Appendix D] The paper should state clearly which experiments use the manually refined DoTA annotations (DoTA_ref) and which use the original DoTA. The text says the refined set is used for comparisons between Video ViT models and the original set is used for comparisons with specialized methods, but the tables do not consistently indicate which set was used, and this affects the comparability of numbers across tables.
  4. [Fig. 3] The DAPT gains in Fig. 3 are reported as single-run differences without error bars or significance testing. For small models the point differences are +2.9 and +3.9 percentage points, but it would be helpful to know whether these exceed run-to-run variance, especially since the fine-tuning recipe involves random sampling of 50K examples per epoch.
  5. [Table 2 footnote] The footnote says ISCRTAD's parameter count, GPU memory, and FPS are 'optimistic estimates using publicly available components.' This is vague; the authors should state exactly which components were measured and which were estimated, and why FPS for prior methods were not measured under the same hardware conditions as their own models.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark study with held-out evaluation; prior-method numbers are external, and self-citations are not load-bearing.

full rationale

This is an empirical benchmark paper rather than a derivation, and I find no circular step. The encoder-only pipeline (Section 3, Appendix A) is a standard fine-tuning recipe: off-the-shelf Video ViT checkpoints plus a linear head, trained with cross-entropy on fixed benchmark splits, with checkpoints selected by validation AUCROC/AUCMCC (Appendix A: 'We report AUC ROC using checkpoints with the highest AUC ROC on the validation set'). The headline comparisons in Tab. 2 are against published numbers from prior specialized methods (marked '†From prior work') on the same benchmark splits; those numbers are inputs to the comparison, not quantities fitted or predicted by this paper's model. Tab. 3 compares external checkpoints from other groups (VideoMAE, MVD, VideoMAE2, UMT, SMILE, InternVideo2) with stated pretraining objectives; the conclusion about MVM is an empirical observation on those checkpoints, and although the comparison is confounded by pretraining data scale and distillation stages, a confound is not a circular reduction. Refs. [18, 19, 47] include the authors' prior work on segmentation, but they are used only as motivational context for studying simple encoders; the TAD result does not depend on those papers' conclusions. No equation in the paper is defined in terms of the quantity it claims to predict, and no fitted parameter is renamed as a prediction. The fragility of beating prior numbers without re-running them under a common protocol is a validity risk, not circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper is empirical and introduces no new theoretical entities or derived constants. Its conclusions rest on benchmark assumptions, uncontrolled comparisons between public checkpoints, and the comparability of prior-work baselines; these are listed as axioms and red flags.

free parameters (2)
  • Fine-tuning recipe (batch size 56, LR 5e-4 for Base/Large, layer decay 0.6, 50 epochs) = 56, 5e-4, 0.6, 50
    Taken from the VideoMAE HMDB51 recipe rather than tuned per TAD dataset; the central comparison depends on this recipe being reasonable for all models.
  • DAPT schedule (batch size 800, 1M samples per epoch, 12 epochs, mask ratio 0.75) = 800, 1M, 12, 0.75
    Chosen as a downscaled VideoMAE pretraining recipe; the DAPT gains in Fig. 3 depend on these settings.
assumptions (4)
  • domain assumption AUCROC on DoTA and DADA-2000 is a valid proxy for traffic anomaly detection quality.
    The paper's main ranking of models rests on these benchmark metrics; AUCMCC is reported for some comparisons but is not used everywhere.
  • domain assumption Publicly released pretrained checkpoints differ mainly in pretraining objective, so TAD rankings can be attributed to supervision type.
    Tab. 3 compares off-the-shelf models that also differ in pretraining data scale, distillation stages, and architecture details; this is not a controlled ablation.
  • domain assumption Prior-work results for TTHF, PromptTAD, and ISCRTAD are directly comparable to the authors' runs.
    Tab. 2 uses literature-reported AUC and efficiency numbers for these methods rather than re-running them under the same protocol.
  • domain assumption DAPT on normal driving videos improves TAD without anomalous examples.
    Supported by Tab. 4, but assumes BDD100K is sufficiently close to the TAD evaluation domain and that the reconstruction objective is the right auxiliary task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simplifying Traffic Anomaly Detection with Video Foundation Models." pith.science (2026). https://pith.science/paper/D6JKZLVB

@misc{pith2026250709338,
  author       = {Pith},
  title        = {Pith review of: Simplifying Traffic Anomaly Detection with Video Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D6JKZLVB}},
  note         = {Machine review of arXiv:2507.09338}
}
read the original abstract

Recent methods for ego-centric Traffic Anomaly Detection (TAD) often rely on complex multi-stage or multi-representation fusion architectures, yet it remains unclear whether such complexity is necessary. Recent findings in visual perception suggest that foundation models, enabled by advanced pre-training, allow simple yet flexible architectures to outperform specialized designs. Therefore, in this work, we investigate an architecturally simple encoder-only approach using plain Video Vision Transformers (Video ViTs) and study how pre-training enables strong TAD performance. We find that: (i) advanced pre-training enables simple encoder-only models to match or even surpass the performance of specialized state-of-the-art TAD methods, while also being significantly more efficient; (ii) although weakly- and fully-supervised pre-training are advantageous on standard benchmarks, we find them less effective for TAD. Instead, self-supervised Masked Video Modeling (MVM) provides the strongest signal; and (iii) Domain-Adaptive Pre-Training (DAPT) on unlabeled driving videos further improves downstream performance, without requiring anomalous examples. Our findings highlight the importance of pre-training and show that effective, efficient, and scalable TAD models can be built with minimal architectural complexity. We release our code, domain-adapted encoders, and fine-tuned models to support future work: https://github.com/tue-mps/simple-tad.

Figures

Figures reproduced from arXiv: 2507.09338 by the authors.

Figure 1
Figure 1. Traffic Anomaly Detection (TAD) performance on DADA-2000 [10]. Simple encoder-only models with advanced pre-training (blue) are faster and more accurate than recent multi￾component architectures (orange and red), see Tabs. 2 and 3. on specialized, complex architectures illustrated in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Types of model architectures for TAD: simple encoder [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. DAPT scaling across different model sizes. Smaller models benefit more. S: Small, B: Base, L: Large variants of the Video ViT. 4.4. Domain-Adaptive Pre-Training (DAPT) Larger ViFMs can be pre-trained on a larger scale and, as a result, exhibit better out-of-the-box generalization across domains, while smaller models have shown to benefit less from longer pre-training due to their limited capacity and faster saturati… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative examples for the effect of DAPT. Predicted anomaly-scores of VideoMAE with and without DAPT (top: Video ViT-Small, bottom: Video ViT-Base). centric driving mixed with anomalies. This setup allows us to evaluate whether the observed improvements stem from do…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Asleep at the Wheel: JEPA's Limitations in Evaluating Novel Driving Data

    cs.CV 2026-08 conditional novelty 5.0 of 10

    A JEPA novelty score that appears to triage driving clips across datasets is actually detecting dataset provenance, and fails on a same-dataset benchmark where a supervised probe on the same embeddings succeeds.

Reference graph

Works this paper leans on

57 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    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. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 6836–6846,

  2. [2]

    Layer normalization

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hin- ton. Layer normalization. arXiv preprint arXiv:1607.06450,

    arXiv

  3. [3]

    A short note on the kinetics-700 human action dataset

    Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zis- serman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 5

    arXiv 1907

  4. [4]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on ma- chine learning, pages 1597–1607. PmLR, 2020. 3

    2020

  5. [5]

    The matthews cor- relation coefficient (mcc) should replace the roc auc as the standard metric for assessing binary classification

    Davide Chicco and Giuseppe Jurman. The matthews cor- relation coefficient (mcc) should replace the roc auc as the standard metric for assessing binary classification. BioData Mining, 16(1):4, 2023. 4

    work page 2023

  6. [6]

    Scaling egocentric vision: The epic-kitchens dataset

    Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Scaling egocentric vision: The epic-kitchens dataset. In Proceedings of the European conference on computer vision (ECCV), pages 720–736, 2018. 1

    work page 2018

  7. [7]

    Flashattention: Fast and memory-efficient exact attention with io-awareness

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christo- pher R ´e. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344–16359, 2022. 3

    work page 2022

  8. [8]

    Cyclecrash: A dataset of bicycle collision videos for col- lision prediction and analysis

    Nishq Poorav Desai, Ali Etemad, and Michael Greenspan. Cyclecrash: A dataset of bicycle collision videos for col- lision prediction and analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2025. 1, 2, 3, 5, 6, 12

    work page 2025

Show all 57 references
  1. [9]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint a...

    2010 arXiv

  2. [10]

    Dada: Driver attention prediction in driving accident scenarios

    Jianwu Fang, Dingxin Yan, Jiahuan Qiao, Jianru Xue, and Hongkai Yu. Dada: Driver attention prediction in driving accident scenarios. IEEE transactions on intelligent trans- portation systems, 23(6):4959–4971, 2021. 1, 2, 3, 5, 6, 7, 13, 14, 15

    2021

  3. [11]

    Cognitive accident prediction in driving scenes: A multimodality benchmark

    Jianwu Fang, Lei-Lei Li, Kuan Yang, Zhedong Zheng, Jianru Xue, and Tat-Seng Chua. Cognitive accident prediction in driving scenes: A multimodality benchmark. CoRR, abs/2212.09381, 2022. 5, 7, 12

    2022 arXiv

  4. [12]

    The” something something” video database for learning and evaluating visual common sense

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michal- ski, 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. In ...

    2017

  5. [13]

    Don’t stop pretraining: Adapt language models to domains and tasks

    Suchin Gururangan, Ana Marasovi ´c, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. Don’t stop pretraining: Adapt language models to domains and tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics . Associat...

    2020

  6. [14]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16000– 16009, 2022. 3

    2022

  7. [15]

    Mgmae: Motion guided masking for video masked autoencoding

    Bingkun Huang, Zhiyu Zhao, Guozhen Zhang, Yu Qiao, and Limin Wang. Mgmae: Motion guided masking for video masked autoencoding. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision , pages 13493– 13504, 2023. 4, 13

    2023

  8. [16]

    An enhanced traffic in- cident detection using factor analysis and weighted random forest algorithm

    P Rajesh Kanna, S Vanithamani, P Karunakaran, P Pandi- araja, N Tamilarasi, and P Nithin. An enhanced traffic in- cident detection using factor analysis and weighted random forest algorithm. In 2024 International Conference on IoT Based Control Networks and Intelligent Systems...

    2024

  9. [17]

    The kinetics hu- man action video dataset

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

    arXiv

  10. [18]

    First Place Solution to the ECCV 2024 BRA VO Challenge: Eval- uating Robustness of Vision Foundation Models for Seman- tic Segmentation

    Tommie Kerssies, Daan de Geus, and Gijs Dubbelman. First Place Solution to the ECCV 2024 BRA VO Challenge: Eval- uating Robustness of Vision Foundation Models for Seman- tic Segmentation. arXiv preprint arXiv:2409.17208 , 2024. 2

    2024 arXiv

  11. [19]

    Your vit is secretly an image segmentation model

    Tommie Kerssies, Niccolo Cavagnero, Alexander Hermans, Narges Norouzi, Giuseppe Averta, Bastian Leibe, Gijs Dubbelman, and Daan de Geus. Your vit is secretly an image segmentation model. In Proceedings of the Computer Vision and Pattern Recognition Conference , pages 25303–25313,

  12. [20]

    Crash to not crash: Learn to identify dangerous vehicles using a simulator

    Hoon Kim, Kangwook Lee, Gyeongjo Hwang, and Changho Suh. Crash to not crash: Learn to identify dangerous vehicles using a simulator. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 978–985, 2019. 3

    2019

  13. [21]

    Hmdb: a large video database for human motion recognition

    Hildegard Kuehne, Hueihan Jhuang, Est ´ıbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 Inter- national conference on computer vision , pages 2556–2563. IEEE, 2011. 12

    2011

  14. [22]

    Unmasked teacher: Towards training-efficient video foundation models

    Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the IEEE/CVF international conference on computer vi- sion, pages 19948–19960, 2023. 3, 4, 5, 7

    2023

  15. [23]

    Videomamba: State space model for efficient video understanding

    Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pages 237–255. Springer, 2024. 3

    2024

  16. [24]

    Text-driven traffic anomaly detection with temporal high- frequency modeling in driving videos

    Rongqin Liang, Yuanman Li, Jiantao Zhou, and Xia Li. Text-driven traffic anomaly detection with temporal high- frequency modeling in driving videos. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2, 3, 4, 6, 14

    2024

  17. [25]

    An interaction-scene collaborative representation framework for detecting traffic anomalies in driving videos

    Rongqin Liang, Yuanman Li, Zhenyu Wu, and Xia Li. An interaction-scene collaborative representation framework for detecting traffic anomalies in driving videos. IEEE Transac- tions on Intelligent Transportation Systems, 2025. 1, 3, 4, 6, 13, 14

    2025

  18. [26]

    Fu- ture frame prediction for anomaly detection–a new baseline

    Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Fu- ture frame prediction for anomaly detection–a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536–6545, 2018. 3

    2018

  19. [27]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 11976–11986,

  20. [28]

    Video swin transformer

    Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202–3211, 2022. 3

    2022

  21. [29]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 12

    2017 arXiv

  22. [30]

    Remembering history with convolutional lstm for anomaly detection

    Weixin Luo, Wen Liu, and Shenghua Gao. Remembering history with convolutional lstm for anomaly detection. In 2017 IEEE International conference on multimedia and expo (ICME), pages 439–444. IEEE, 2017. 2

    2017

  23. [31]

    Foundation models for video understanding: A survey

    Neelu Madan, Andreas Møgelmose, Rajat Modi, Yogesh S Rawat, and Thomas B Moeslund. Foundation models for video understanding: A survey. Authorea Preprints, 2024. 3

    2024

  24. [32]

    Howto100m: Learning a text-video embedding by watching 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 watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision , pages...

    2019

  25. [33]

    End-to-end learning of visual representations from uncurated instruc- tional videos

    Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instruc- tional videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages...

    2020

  26. [34]

    Viorica P ˘atr˘aucean, Xu Owen He, Joseph Heyward, Chuhan Zhang, Mehdi S. M. Sajjadi, George-Cristian Muraru, Artem Zholus, Mahdi Karami, Ross Goroshin, Yutian Chen, Simon Osindero, Jo˜ao Carreira, and Razvan Pascanu. Trecvit: A re- current video transformer. arXiv preprint ar...

  27. [35]

    Prompttad: Object-prompt enhanced traffic anomaly detection

    Hao Qiu, Xiaobo Yang, and Xiaojin Gong. Prompttad: Object-prompt enhanced traffic anomaly detection. IEEE Robotics and Automation Letters , 2025. 1, 2, 3, 4, 6, 13, 14

    2025

  28. [36]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  29. [37]

    Memory-augmented online video anomaly detection

    Leonardo Rossi, Vittorio Bernuzzi, Tomaso Fontanini, Mas- simo Bertozzi, and Andrea Prati. Memory-augmented online video anomaly detection. In ICASSP 2024-2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pages 6590–6594. IEEE, 2024. 1...

    2024

  30. [38]

    Sigma: Sinkhorn-guided masked video mod- eling

    Mohammadreza Salehi, Michael Dorkenwald, Fida Mo- hammad Thoker, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Sigma: Sinkhorn-guided masked video mod- eling. In European Conference on Computer Vision , pages 293–312. Springer, 2024. 4, 12, 13

    2024

  31. [39]

    Learning to predict collision risk from sim- ulated video data

    Tim J Schoonbeek, Fabrizio J Piva, Hamid R Abdolhay, and Gijs Dubbelman. Learning to predict collision risk from sim- ulated video data. In 2022 IEEE Intelligent Vehicles Sympo- sium (IV), pages 943–951. IEEE, 2022. 3

    2022

  32. [40]

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

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1

    2012 arXiv

  33. [41]

    Masked motion encoding for self-supervised video representation learning

    Xinyu Sun, Peihao Chen, Liangwei Chen, Changhao Li, Thomas H Li, Mingkui Tan, and Chuang Gan. Masked motion encoding for self-supervised video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2235–2245,

  34. [42]

    Deep learning applied to road accident detection with transfer learning and synthetic im- ages

    Tiago Tamagusko, Matheus Gomes Correia, Minh Anh Huynh, and Adelino Ferreira. Deep learning applied to road accident detection with transfer learning and synthetic im- ages. Transportation research procedia, 64:90–97, 2022. 4

    2022

  35. [43]

    Smile: Infusing spatial and motion se- mantics in masked video learning

    Fida Mohammad Thoker, Letian Jiang, Chen Zhao, and Bernard Ghanem. Smile: Infusing spatial and motion se- mantics in masked video learning. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8438–8449, 2025. 3, 4, 5, 7

    2025

  36. [44]

    Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35:10078–10093, 2022. 2, 3, 4, 5, 7, 12, 13, 14 10

    2022

  37. [45]

    A closer look at spatiotemporal convolutions for action recognition

    Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recogni- tion, pages 6450–6459, 2018. 3, 5, 6, 12

    2018

  38. [46]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3

    2017

  39. [47]

    The BRA VO Semantic Segmentation Challenge Results in UNCV2024

    Tuan-Hung Vu, Eduardo Valle, Andrei Bursuc, Tommie Kerssies, Daan de Geus, Gijs Dubbelman, Long Qian, Bingke Zhu, Yingying Chen, Ming Tang, Jinqiao Wang, Tom´aˇs V oj´ıˇr, Jan ˇSochman, Jiˇr´ı Matas, Michael Smith, Frank Ferrie, Shamik Basu, Christos Sakaridis, and Luc Van Goo...

    2024

  40. [48]

    Rs2g: Data-driven scene-graph extraction and embedding for ro- bust autonomous perception and scenario understanding

    Junyao Wang, Arnav Vaibhav Malawade, Junhong Zhou, Shih-Yuan Yu, and Mohammad Abdullah Al Faruque. Rs2g: Data-driven scene-graph extraction and embedding for ro- bust autonomous perception and scenario understanding. In Proceedings of the IEEE/CVF Winter Conference on Appli- c...

    2024

  41. [49]

    Abnormal event detection in videos using hy- brid spatio-temporal autoencoder

    Lin Wang, Fuqiang Zhou, Zuoxin Li, Wangxia Zuo, and Haishu Tan. Abnormal event detection in videos using hy- brid spatio-temporal autoencoder. In 2018 25th IEEE In- ternational Conference on Image Processing (ICIP) , pages 2276–2280. IEEE, 2018. 2

    2018

  42. [50]

    Videomae v2: Scaling video masked autoencoders with dual masking

    Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi- nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 14549–14560, 2023. 4, 5, 7

    2023

  43. [51]

    Masked video distillation: Rethinking masked feature mod- eling for self-supervised video representation learning

    Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature mod- eling for self-supervised video representation learning. In Proceedings of the IEEE/CVF conference on computer vi- ...

    2023

  44. [52]

    Internvideo2: Scaling foundation models for mul- timodal video understanding

    Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for mul- timodal video understanding. In European Conference on Computer Vision, pages 396–416. Springer, 2024. 4, 5, 7, 12

    2024

  45. [53]

    Recurring the transformer for video action recognition

    Jiewen Yang, Xingbo Dong, Liujun Liu, Chao Zhang, Jia- jun Shen, and Dahai Yu. Recurring the transformer for video action recognition. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 14063–14073, 2022. 3

    2022

  46. [54]

    Dota: unsupervised detec- tion of traffic anomaly in driving videos

    Yu Yao, Xizi Wang, Mingze Xu, Zelin Pu, Yuchen Wang, Ella Atkins, and David Crandall. Dota: unsupervised detec- tion of traffic anomaly in driving videos. IEEE transactions on pattern analysis and machine intelligence, 2022. 1, 2, 3, 4, 5, 6, 12, 13, 14

    2022

  47. [55]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar- rell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition ...

    2020

  48. [56]

    Scaling vision transformers

    Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lu- cas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104–12113, 2022. 3, 7

    2022

  49. [57]

    Spatio-temporal feature encoding for traffic accident detection in vanet environment

    Zhili Zhou, Xiaohua Dong, Zhetao Li, Keping Yu, Chun Ding, and Yimin Yang. Spatio-temporal feature encoding for traffic accident detection in vanet environment. IEEE Transactions on Intelligent Transportation Systems, 23(10): 19772–19781, 2022. 1, 4, 14 11 Appendix In this app...

    2022

Pith tools

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