Pith. sign in

REVIEW 4 major objections 8 minor 1 cited by

Driver Assistance System Based on Multimodal Data Hazard Detection

T0 review · 4 major / 8 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper claims that fusing road video, driver face video, and driver audio in a single end-to-end attention-based model recognizes dangerous driving states with 96.875% accuracy, beating every one- or two-modality baseline on its…

desk verdict Standard tri-modal fusion stack meets a private 32-clip dataset; the headline accuracy is a one-sample difference and the central claim is not established. read the letter →

arxiv 2502.03005 v1 pith:XP6H32SM submitted 2025-02-05 cs.CV cs.LG

Long Zhouxiang , Ovanes Petrosian This is my paper · ORCID
classification cs.CVcs.LG
keywords multimodalfusiondangerousdrivingdetectiondriverassistanceattentionmechanismroadvideoanalysisfacialaudiosimulateddataset
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

Most hazard-detection systems for driving look only at road video, which misses rare events and gives no information about the person behind the wheel. This paper tries to establish that adding two more data streams—a camera on the driver's face and a microphone for the driver's voice—and fusing all three in a single end-to-end attention-based model improves dangerous-driving recognition. On the paper's own simulator dataset, the tri-modal model reaches 96.875% accuracy on a safe-versus-dangerous clip classification task, beating every unimodal and bimodal combination it was tested against. The paper's interpretation is that the three modalities carry complementary cues: facial expression and gaze signal attention or fatigue, voice and breathing signal emotion or drowsiness, and road video signals scene-level risk. The evidence is generated in a driving simulator, so the claim is an existence proof for the fusion architecture rather than a demonstration on real roads.

What carries the argument

The central mechanism is the attention-based intermediate fusion module. For any two modality feature maps, the model computes a pairwise similarity matrix $A_{ij} = \operatorname{softmax}\left( \frac{\Phi_i W_q W_k^T \Phi_j^T}{\sqrt{d}} \right)$ and uses those scores to reweight features of one modality by their consistency with another. This is not concatenation or late averaging; each branch keeps its own learned representation while sharing information through attention scores. Supporting this are the visual branches' depthwise separable convolutions plus 1D temporal convolutional blocks, the audio branch's Mel-frequency cepstral coefficient features with 1D convolutional blocks, and the road branch's channel shuffling and channel/spatial attention.

What would settle it

Run the trained tri-modal model on a held-out set of real road, driver-face, and microphone recordings with human-verified hazard labels; if its accuracy falls to the level of the road-video-only baseline, the claimed fusion advantage is an artifact of the simulator. A simpler probe is to flip the '02' dangerous tags on a random subset of simulator clips and retrain; if accuracy tracks the flipped filenames rather than the actual content, the classifier has learned a shortcut.

Watch

Extended reading notes

Core claim

The paper reports one central result: a three-branch neural network that takes raw road-condition video, driver face video, and driver audio, and fuses them with pairwise attention at an intermediate layer, classifies dangerous driving states better than any one modality or any two-modality combination. On the constructed simulator dataset the full A-V-R model scores 96.875% accuracy with cross-entropy loss 0.0122, while the best bimodal model reaches 93.75%, the best unimodal model reaches 87.5%, and the replicated road-video-only anomaly detectors reach at most 92.45%. The paper interprets this as evidence that the modalities are complementary and synergistic: facial cues capture attention and fatigue, audio captures emotion and yawning, road video captures lane deviation and proximity risk. The pairwise attention fusion keeps features that agree across modalities while suppressing inconsistent ones, which the authors argue is why the full model outperforms every subset.

Load-bearing premise

The result depends on the assumption that the simulator's 'dangerous' clips, which the authors generated and labeled by filename, contain the same hazard-relevant patterns that occur in real driving; if the model is instead exploiting simulation-specific or filename-specific cues, the 96.875% figure will not survive contact with real road data.

Editorial extensions

If this is right

  • If the reported result holds, tri-modal attention fusion beats every single- and paired-modality configuration on the paper's task, making a concrete case for going beyond road-video-only hazard detection.
  • The end-to-end design removes the need for separately pre-trained feature extractors, which the paper argues reduces pipeline uncertainty in real-time assisted-driving systems.
  • Because fusion uses attention scores rather than concatenated raw features, the model is claimed to tolerate incomplete or noisy modalities, a property the data-augmentation experiments probe.
  • On this dataset the tri-modal model also outperforms replicated road-video anomaly detectors, indicating that the advantage is not simply a larger parameter count but the combination of modalities.

Reading between the lines

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

  • A test the authors list only as future work is cross-dataset transfer: training on this simulator data and evaluating on real dashcam and driver-monitoring recordings would reveal whether the attention fusion learns hazard semantics or simulator appearance.
  • The binary clip-level accuracy leaves open the harder problem of localizing when in the clip a hazard starts; extending the same architecture to frame-level or temporal-segment prediction would be more useful for issuing timely alerts.
  • The pairwise attention fusion is modality-agnostic, so it could be extended to a fourth stream such as vehicle telemetry (speed, acceleration, lane offset), which the paper mentions as future work but does not test.
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 / 8 minor

Summary. The manuscript proposes a driver assistance hazard detection system that fuses three modalities—road condition video, driver face video, and driver audio—through an attention-based intermediate fusion mechanism that operates end-to-end. The authors introduce a new dataset recorded in a driving simulator (AirSim and Need for Speed), with safe/dangerous labels encoded in file names. In experiments, the tri-modal model is reported to reach 96.875% accuracy on dangerous driving state recognition, outperforming unimodal and bimodal ablations (Table 2) and two video anomaly detection baselines (Table 3). The conclusion is that tri-modal attention fusion improves hazard detection and that the proposed fusion strategy is effective.

Significance. If the reported 96.875% accuracy were properly established, the work would be a useful contribution: it addresses an important safety application, proposes an end-to-end trainable architecture that avoids separate offline feature extraction, and contributes a tri-modal driving dataset, which is genuinely scarce. The design of modality-specific branches with an intermediate attention-based fusion is reasonable, and Table 1 provides a clear architecture specification. However, the manuscript provides no credible statistical evidence for the central claim: the evaluation is based on a very small test set (likely 32 clips), self-labeled simulated data without external verification, and baselines that are not comparable to the proposed model. As it stands, the paper does not substantiate the claimed superiority of tri-modal fusion or the specific attention mechanism.

major comments (4)
  1. [Section 6, Table 2] Every reported accuracy is an exact multiple of 1/32 (96.875 = 31/32, 93.75 = 30/32, 90.625 = 29/32, 87.5 = 28/32, 81.25 = 26/32, 78.125 = 25/32, 46.875 = 15/32). This strongly indicates a 32-clip test set, yet the manuscript never states the test-set size, train/test split, number of repetitions, or confidence intervals. The tri-modal model's edge over the best paired model is 31/32 versus 30/32, i.e., one clip. Without per-clip predictions or a paired significance test, the numbers in Table 2 do not establish that adding the audio and face modalities improves hazard detection. The central accuracy claim therefore rests on a one-sample difference at an unknown split.
  2. [Section 4.2] The dataset labels are generated by the authors (safe/dangerous denoted by the second filename field '01'/'02') and no independent annotation, validation, or inter-annotator agreement is reported. Since the authors also created the simulator scenarios, the classifier could be learning to separate easy simulation-specific cues (e.g., scripted obstacle appearances or synthetic audio) rather than general hazard patterns. The absence of any external benchmark or real-world evaluation means the reported 96.875% accuracy has unknown external validity. This is a load-bearing limitation for a paper whose contribution is a driving safety system.
  3. [Section 6, Table 3] The comparison to AstNet and WideBranchNet is not an ablation of the proposed method: those baselines are evaluated on road-video data only ('We used our own road condition video data to evaluate these models'), whereas the proposed model receives three modalities. Thus the reported accuracy gap cannot isolate the contribution of the fusion architecture; it could be due to additional information channels or to differences in training protocols, dataset composition, and hyperparameters. Moreover, Table 3 again reports no test-set size or variance, so the differences (e.g., 96.875 vs 92.45) are not statistically interpretable.
  4. [Section 5.3 and Section 6] The claim that the attention-based intermediate fusion mechanism is responsible for the improvement is not supported by any comparison to alternative fusion strategies, such as early concatenation, late fusion, or simple feature addition. Since the only experimental comparisons vary the number of input modalities while keeping the fusion module fixed, the paper does not demonstrate that the proposed attention mechanism outperforms simpler fusion baselines. Given that the central contribution of the paper is this fusion strategy, an ablation that varies the fusion module is essential for the stated claims.
minor comments (8)
  1. [Section 3 and Section 6] There are multiple typos and grammatical errors, e.g., 'wiil' in Section 3 and 'triflod-multimodal' in Section 6, which should be corrected.
  2. [Section 4.1 and Section 5.1] Section 4.1 states that data are saved in 3-second units, while Section 5.1 uses a fixed audio length of 3.6 seconds; the discrepancy is unexplained.
  3. [Section 6] The text refers to 'Table 6.2' when describing Table 3; the table reference is incorrect.
  4. [Section 5.3 and Section 6] The manuscript uses 'two-dimensional' and 'three-dimensional' models to mean bimodal and trimodal inputs, which conflates modality count with spatial dimensionality and makes the prose confusing.
  5. [Section 2] Several sentences are incomplete or have broken references, e.g., 'data5 collected 1,500 video clips...' lacks a proper citation and reads as a fragment.
  6. [Equations (1) and (2)] The symbol d is defined as the dimensionality of a latent space in Eq. (1) but as the number of columns of the matrix in the following sentence; this notation should be made consistent.
  7. [Table 1] The audio branch uses 'MPool1D' for max-pooling, inconsistent with 'MaxPool1d' used earlier in the same table; please unify the notation.
  8. [Abstract and Section 5.2] The abstract claims 'end-to-end learning without separate feature extraction,' but Section 5.2 relies on MTCNN for face detection and cropping; the paper should clarify that the face detection step is not learned end-to-end.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the tri-modal classifier is empirically trained and tested, with no derivation that reduces to its own inputs.

full rationale

Reviewing the claimed derivation chain from Section 3 (model) through Section 6 (results), there is no equation-level or citation-level circularity. The attention fusion formula (Eq. 2) is a standard operation; no parameter is fitted to the test-set outcome and renamed as a prediction. The labels in the custom dataset are derived from simulator event/filename annotation, but the model inputs are raw/processed video and audio tensors, not the labels, so the accuracy numbers are empirical measurements rather than consequences of the annotation scheme. The manuscript contains no self-citations that carry the argument, and the 32-clip/one-sample statistical concern raised by the skeptic is a validity/reliability issue, not an instance of the paper's central claim reducing by construction to its inputs. Therefore the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the validity of the authors' simulator dataset, the labeling scheme, and the fixed preprocessing extractors. The model itself is a standard supervised classifier, so most assumptions are domain assumptions about data representativeness rather than mathematical axioms. No new physical entities are introduced.

free parameters (6)
  • All trainable network weights (three branches and fusion modules) = not reported
    The 96.875% accuracy claim is produced by these fitted parameters; the paper gives no parameter count, optimizer configuration, or run-to-run variance.
  • Audio clip length = 3.6 s
    Chosen by hand in Section 5.1; classification depends on this fixed window.
  • Sampled frames per video = 15
    Chosen by hand in Section 5.2; temporal coverage depends on this choice.
  • MFCC parameters = not specified
    MFCC coefficients are used as audio input; number of coefficients, window, and hop length are not reported.
  • Fusion attention latent dimension d = not reported
    Equation (2) scales attention by sqrt(d); d is never stated.
  • Optimizer hyperparameters = not reported (SGD with momentum, dynamic LR, dropout)
    Section 5.3 mentions these but gives no values.
assumptions (5)
  • domain assumption The AirSim and Need for Speed simulator footage is a valid proxy for real driving hazards.
    Section 4 states the dataset is generated in a simulator and game; the central generalization claim rests on this proxy.
  • domain assumption Filename-based labels ('02' = dangerous) are accurate and sufficient.
    Section 4.2: annotation is encoded in the filename; no human verification or inter-annotator agreement is described.
  • domain assumption MTCNN face detection and MFCC preprocessing capture all task-relevant information.
    Sections 5.1-5.2 introduce these fixed extractors; the model cannot see information they discard.
  • ad hoc to paper A model that is more complex generalizes better.
    Section 6 and Conclusion infer stronger generalization from higher complexity, which is not logically implied.
  • domain assumption Attention-based intermediate fusion is better than early or late fusion for this task.
    The paper motivates this from [19] but provides no ablation comparing fusion strategies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Driver Assistance System Based on Multimodal Data Hazard Detection." pith.science (2026). https://pith.science/paper/XP6H32SM

@misc{pith2026250203005,
  author       = {Pith},
  title        = {Pith review of: Driver Assistance System Based on Multimodal Data Hazard Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XP6H32SM}},
  note         = {Machine review of arXiv:2502.03005}
}
read the original abstract

Autonomous driving technology has advanced significantly, yet detecting driving anomalies remains a major challenge due to the long-tailed distribution of driving events. Existing methods primarily rely on single-modal road condition video data, which limits their ability to capture rare and unpredictable driving incidents. This paper proposes a multimodal driver assistance detection system that integrates road condition video, driver facial video, and audio data to enhance incident recognition accuracy. Our model employs an attention-based intermediate fusion strategy, enabling end-to-end learning without separate feature extraction. To support this approach, we develop a new three-modality dataset using a driving simulator. Experimental results demonstrate that our method effectively captures cross-modal correlations, reducing misjudgments and improving driving safety.

Figures

Figures reproduced from arXiv: 2502.03005 by the authors.

Figure 1
Figure 1. End-to-end recognitions framework Currently, there is a lack of publicly available datasets containing road condition video data, driver facial video data, and driver audio data simultaneously. Consequently, this paper establishes a three-modality dataset by using the Arisim driving simulation project and Logitech driving simulation devices to conduct simulated driving while recording road condition video data, driv… view at source ↗
Figure 2
Figure 2. Tri-modal recognitions framework 8 [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Traing Process To comprehensively evaluate the overall performance of the model, we meticulously design three ablation experiments. These experiments aim to compare the specific impacts of different modal inputs (including 13 [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Traing Process 6 Result Comprehensive experimental evaluations conducted on the constructed dataset indicate that the proposed three-dimensional model is capable of achieving outstanding performance in dangerous driving state recognition. The results of the comparative…

Discussion (0). Continue with ORCID 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. RoadFed: A Multimodal Federated Learning System for Improving Road Safety

    cs.CE 2025-02 reject novelty 4.0 of 10

    A multimodal federated learning system with quantization and local differential privacy is reported to detect road hazards at 96.42% accuracy with 0.035 s latency and up to 1000x lower communication cost than baselines.

Reference graph

Works this paper leans on

35 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    Crash fatality rates after recreational marijuana legalization in washington and colorado.American journal of public health, 107(8):1329–1331, 2017

    Jayson D Aydelotte, Lawrence H Brown, Kevin M Luftman, Alexandra L Mardock, Pedro GR Teixeira, Ben Coopwood, and Carlos VR Brown. Crash fatality rates after recreational marijuana legalization in washington and colorado.American journal of public health, 107(8):1329–1331, 2017

    work page 2017

  2. [2]

    Hybrid conditional random field based camera-lidar fusion for road detection.Information Sciences, 432:543–558, 2018

    Liang Xiao, Ruili Wang, Bin Dai, Yuqiang Fang, Daxue Liu, and Tao Wu. Hybrid conditional random field based camera-lidar fusion for road detection.Information Sciences, 432:543–558, 2018

    work page 2018

  3. [3]

    Rbnet: A deep neural network for unified road and road boundary detection

    Zhe Chen and Zijing Chen. Rbnet: A deep neural network for unified road and road boundary detection. In Neural Information Processing: 24th International Conference, ICONIP 2017, Guangzhou, China, November 14-18, 2017, Proceedings, Part I 24, pages 677–687. Springer, 2017

    work page 2017

  4. [4]

    Multinet: Real-time joint semantic reasoning for autonomous driving

    Marvin Teichmann, Michael Weber, Marius Zoellner, Roberto Cipolla, and Raquel Urtasun. Multinet: Real-time joint semantic reasoning for autonomous driving. In2018 IEEE intelligent vehicles symposium (IV), pages 1013–1020. IEEE, 2018

    work page 2018

  5. [5]

    Transfer learning based hybrid 2d-3d cnn for traffic sign recognition and semantic road detection applied in advanced driver assistance systems

    Khaled Bayoudh, Fayçal Hamdaoui, and Abdellatif Mtibaa. Transfer learning based hybrid 2d-3d cnn for traffic sign recognition and semantic road detection applied in advanced driver assistance systems. Applied Intelligence, 51(1):124–142, 2021

    work page 2021

  6. [6]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105–6114. PMLR, 2019

    2019

  7. [7]

    Long-term Multi-granularity Deep Framework for Driver Drowsiness Detection

    Jie Lyu, Zejian Yuan, and Dapeng Chen. Long-term multi-granularity deep framework for driver drowsiness detection. arXiv preprint arXiv:1801.02325, 2018

    work page Pith review arXiv 2018

  8. [8]

    Real time detection system of driver drowsiness based on representation learning using deep neural networks.Journal of Intelligent & Fuzzy Systems, 36(3):1977– 1985, 2019

    Vineetha Vijayan and Elizabeth Sherly. Real time detection system of driver drowsiness based on representation learning using deep neural networks.Journal of Intelligent & Fuzzy Systems, 36(3):1977– 1985, 2019

    work page 1977

Show all 35 references
  1. [9]

    Driver drowsiness detection system based on feature representation learning using various deep networks

    Sanghyuk Park, Fei Pan, Sunghun Kang, and Chang D Yoo. Driver drowsiness detection system based on feature representation learning using various deep networks. InAsian conference on computer vision, pages 154–164. Springer, 2016

    2016

  2. [10]

    The expression of the emotions in man and animals, by charles darwin, 1999

    Angela Priestley. The expression of the emotions in man and animals, by charles darwin, 1999

    1999

  3. [11]

    Emotion on the road: necessity, acceptance, and feasibility of affective computing in the car.Advances in human-computer interaction, 2010:1–17, 2010

    Florian Eyben, Martin Wöllmer, Tony Poitschke, Björn Schuller, Christoph Blaschke, Berthold Färber, and Nhu Nguyen-Thien. Emotion on the road: necessity, acceptance, and feasibility of affective computing in the car.Advances in human-computer interaction, 2010:1–17, 2010. 19

    2010

  4. [12]

    Emotion and gesture recognition with soft computing tool for drivers assistance system in human centered transportation

    Urvashi Agrawal, Shubhangi Giripunje, and Preeti Bajaj. Emotion and gesture recognition with soft computing tool for drivers assistance system in human centered transportation. In 2013 IEEE International Conference on Systems, Man, and Cybernetics, pages 4612–4616. IEEE, 2013

    2013

  5. [13]

    Performance comparison of facial emotion recognition: a transfer learning-based driver assistance framework for in-vehicle applications

    Goutam Kumar Sahoo, Santos Kumar Das, and Poonam Singh. Performance comparison of facial emotion recognition: a transfer learning-based driver assistance framework for in-vehicle applications. Circuits, Systems, and Signal Processing, 42(7):4292–4319, 2023

    2023

  6. [14]

    A review of automated vehicle systems

    Szilveszter Siroki, Ádám Nyerges, and Viktor Tihanyi. A review of automated vehicle systems. In2019 International IEEE Conference and Workshop in Óbuda on Electrical and Power Engineering (CANDO- EPE), pages 109–114, 2019

    2019

  7. [15]

    Spatio-temporal action graph networks

    Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. Spatio-temporal action graph networks. In2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 2347–2356, 2019

    2019

  8. [16]

    Driver-behavior modeling using on-road driving data: A new application for behavior signal processing.IEEE Signal Processing Magazine, 33(6):14–21, 2016

    Chiyomi Miyajima and Kazuya Takeda. Driver-behavior modeling using on-road driving data: A new application for behavior signal processing.IEEE Signal Processing Magazine, 33(6):14–21, 2016

    2016

  9. [17]

    Nidhi Kalra and Susan M. Paddock. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? Transportation Research Part A: Policy and Practice, 94:182–193, 2016

    2016

  10. [18]

    Virtual assistants and self-driving cars

    Giuseppe Lugano. Virtual assistants and self-driving cars. In 2017 15th International Conference on ITS Telecommunications (ITST), pages 1–5, 2017

    2017

  11. [19]

    Self-attention fusion for audiovisual emotion recognition with incomplete data

    Kateryna Chumachenko, Alexandros Iosifidis, and Moncef Gabbouj. Self-attention fusion for audiovisual emotion recognition with incomplete data. In2022 26th International Conference on Pattern Recognition (ICPR), pages 2822–2828, 2022

    2022

  12. [20]

    Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov

    Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J. Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov. Multimodal transformer for unaligned multimodal language sequences.Proceedings of the conference. Association for Computational Linguistics. Meeting, 2019:6558–6...

    2019

  13. [21]

    Multimodal emotion recognition using cross-modal attention and 1d convolutional neural networks

    Krishnan N and Ankita Patil. Multimodal emotion recognition using cross-modal attention and 1d convolutional neural networks. InInterspeech, 2020

    2020

  14. [22]

    Airsim-w: A simulation environment for wildlife conservation with uavs

    Elizabeth Bondi, Debadeepta Dey, Ashish Kapoor, Jim Piavis, Shital Shah, Fei Fang, Bistra Dilkina, Robert Hannaford, Arvind Iyer, Lucas Joppa, and Milind Tambe. Airsim-w: A simulation environment for wildlife conservation with uavs. InCOMPASS ’18: Proceedings of the 1st ACM SI...

    2018

  15. [23]

    Anomaly detection and localization in crowded scenes

    Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(1):18–32, 2014

    2014

  16. [24]

    Future frame prediction for anomaly detection - a new baseline

    Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - a new baseline. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6536–6545, 2018

    2018

  17. [25]

    Real-world anomaly detection in surveillance videos

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

    2018

  18. [26]

    Anticipating accidents in dashcam videos

    Fu-Hsiang Chan, Yu-Ting Chen, Yu Xiang, and Min Sun. Anticipating accidents in dashcam videos. In Shang-Hong Lai, Vincent Lepetit, Ko Nishino, and Yoichi Sato, editors,Computer Vision – ACCV 2016, pages 136–153, Cham, 2017. Springer International Publishing

    2016

  19. [27]

    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 Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2633–...

    2020

  20. [28]

    Montero, and Fernando Fernández-Martínez

    Cristina Luna-Jiménez, Ricardo Kleinlein, David Griol, Zoraida Callejas, Juan M. Montero, and Fernando Fernández-Martínez. A proposal for multimodal emotion recognition using aural transformers and action units on ravdess dataset.Applied Sciences, 12(1), 2022

    2022

  21. [29]

    Ali Mollahosseini, Behzad Hasani, and Mohammad H. Mahoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild.IEEE Transactions on Affective Computing, 10:18–31, 2017

    2017

  22. [30]

    Introducing the recola multimodal corpus of remote collaborative and affective interactions

    Fabien Ringeval, Andreas Sonderegger, Juergen Sauer, and Denis Lalanne. Introducing the recola multimodal corpus of remote collaborative and affective interactions. In2013 10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1–8, 2013

    2013

  23. [31]

    G. Wang, J. Li, Z. Wu, et al. Efficientface: an efficient deep network with feature enhancement for accurate face detection.Multimedia Systems, 29:2825–2839, 2023

    2023

  24. [32]

    Abdul and Abdulbasit K

    Zrar Kh. Abdul and Abdulbasit K. Al-Talabani. Mel frequency cepstral coefficient and its applications: A review. IEEE Access, 10:122136–122158, 2022

    2022

  25. [33]

    A survey on deep learning for multimodal data fusion

    Jing Gao, Peng Li, Zhikui Chen, and Jianing Zhang. A survey on deep learning for multimodal data fusion. Neural Computation, 32(5):829–864, 2020

    2020

  26. [34]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in neural information processing systems, pages 5998–6008, 2017. 21

    2017

  27. [35]

    Attention-based residual autoencoder for video anomaly detection

    Viet-Tuan Le and Yong-Guk Kim. Attention-based residual autoencoder for video anomaly detection. Applied Intelligence, 53(3):3240–3254, 2023. 22

    2023

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.