Real-time Traffic Accident Anticipation with Feature Reuse

Inpyo Song; Jangwon Lee

REVIEW 3 major objections 3 minor 22 references

Real-time Traffic Accident Anticipation with Feature Reuse

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read RARE shows that reusing one detector's feature maps can anticipate traffic accidents at 73 frames per second, matching or beating prior accuracy on two dashcam benchmarks.

desk verdict RARE is a plausible, clearly written efficiency-focused TAA method whose real novelty is the attention ranking loss, but the paper's central speed and SOTA claims rest on comparisons that are not apples-to-apples. read the letter →

arxiv 2505.17449 v1 pith:BFCROFDZ submitted 2025-05-23 cs.CV

Inpyo Song , Jangwon Lee This is my paper · ORCID

classification cs.CV

keywords trafficaccidentanticipationreal-timeinferencefeaturereuseobjectdetectionattentionrankinglossautonomousdrivingdashcamvideointerpretability

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

RARE is a traffic-accident anticipation model built to show that accurate forecasting does not require heavy auxiliary modules. It reuses the intermediate feature maps of a single pre-trained object detector as the only visual representation, feeding scene-level features into a GRU and object-level features from RoI Align into a multi-head attention module. A new Attention Score Ranking Loss explicitly forces the model to give higher attention to accident-related objects than to irrelevant ones. The paper reports 73.3 FPS (13.6 ms per frame) on an RTX 6000, a 4-8 times speedup over prior systems, while achieving state-of-the-art Average Precision on the DAD and CCD benchmarks. If these numbers hold, accident anticipation can run in real time without giving up accuracy.

What carries the argument

The load-bearing mechanism is the reuse of a single detector's multi-scale embeddings. The same YOLOv10 backbone features drive a GRU for scene dynamics, and RoI Align on the backbone and neck features produces object embeddings that pass through CBAM, are concatenated with box coordinates, and are fused with scene context by multi-head attention. The attention scores $S_n$ are then supervised by the Attention Score Ranking Loss, $L_R = \min(0, \max(S_{na}) + m - \min(S_a))$, a margin-based ranking loss that requires every accident-overlapping box (IoU > 0.5) to be attended to more than every non-accident box. A temporal queue of fused features plus the AdaLEA loss convert these representations into an accident risk score. Everything outside the detector is intentionally small.

What would settle it

Re-implement a baseline such as DSTA or GSC using the same YOLOv10 detector, the same 640x640 preprocessing, and the same RTX 6000 GPU, then measure end-to-end latency including object detection; if it reaches similar FPS and AP, the claimed 4-8x speedup and the attribution of the accuracy gain to RARE's design collapse.

Watch

Extended reading notes

Core claim

The central claim is that a detector-centric pipeline, with no optical flow, tracking, or separate feature extractor, reaches or exceeds prior accuracy while cutting latency by an order of magnitude. Feature reuse is the mechanism: the backbone and neck features of YOLOv10 already carry the scene and object cues needed to anticipate collisions, so RARE only adds a lightweight GRU, RoI Align, CBAM, attention fusion, and a two-layer classifier. The Attention Score Ranking Loss, which penalizes any non-accident object receiving higher attention than any accident-related object, is what the paper credits for both the accuracy gain and the interpretability. The reported results are 62.2% AP on DAD and 99.8% AP on CCD, with mTTA of 2.51 s and 4.03 s respectively, at 73.3 FPS.

Load-bearing premise

The reported 4-8x speedup treats FPS and latency numbers from prior papers, measured on different hardware and with different preprocessing, as directly comparable to RARE's own measurements, and it assumes the accuracy gain comes from RARE's design rather than from the pre-trained YOLOv10 backbone.

Editorial extensions

If this is right

On the DAD and CCD benchmarks, a detector-only pipeline can match or beat specialized accident-anticipation networks in AP while running at 73.3 FPS, well above the 30 FPS real-time threshold.
Because scene and object features come from the same detector, adding accident anticipation to an existing perception stack costs only a GRU, an attention module, and a small classifier, not a parallel feature-extraction network.
The Attention Score Ranking Loss contributes about 5.9 AP points on DAD, so explicitly supervising where the model looks improves both prediction and interpretability.
Dropping either backbone or neck features costs 8.2 or 6.2 AP points, so reusing both scales of the detector's representation is necessary for the reported accuracy.
RARE keeps a higher mTTA than baselines trained to similar AP, suggesting the speed gain does not come from making later, easier predictions.

Reading between the lines

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

Extension: the published speed comparison mixes measurements taken on different hardware and preprocessing conditions, so a strict one-machine re-run of all baselines could shrink the claimed 4-8x gap; the design advantage would be better tested on identical equipment.
Extension: because the backbone choice is not ablated, part of the AP gain may come from the detector rather than from feature reuse or the ranking loss; replacing the detector would isolate this.
Extension: the same reuse-the-detector's-embeddings recipe should transfer to other real-time video tasks such as pedestrian intention prediction or driver monitoring, where a pretrained detector already exists.
Extension: the attention ranking loss assumes ground-truth accident boxes are available at training time; in fully unlabeled settings a proxy such as motion saliency or risk heatmaps would be needed to keep the interpretability benefit.

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.

Desk Editor's Note

read the letter

First, the headline: RARE is a pragmatic, well-written accident anticipation method that reuses detector features and adds an explicit attention ranking loss. That loss is the genuinely new piece; feature reuse from a single detector is a sensible engineering choice, and the ablations show it matters. If you work on TAA or efficient video reasoning, this is worth a look.

What the paper does well: the idea of eliminating optical flow and separate feature extractors by pulling features from YOLOv10 is clearly explained, and the attention ranking loss is a new objective that directly supervises which objects the model attends to. The ablations show the loss helps a lot (5.9 AP on DAD). The writing is direct and the hardware limitations are at least acknowledged.

The soft spots are in the evaluation, and they are not minor. Table 1 mixes FPS values from different GPUs and preprocessing pipelines. The authors take numbers from original papers and only add a measured overhead for GSC; they do not re-run baselines on the RTX 6000. Because YOLOv10 is an extremely efficient detector, the claimed 4-8x speedup could largely be from the detector and hardware, not from the architecture. The AP comparison has the same issue: no baseline uses the same YOLOv10 backbone, so the 1.8-point gain over GSC on DAD could be from the backbone rather than the ranking loss. On top of that, there are no error bars, and the mTTA numbers for UString and DSTA change between Table 1 and Table 2—the paper explains they are from a different training run, but it's easy to confuse and should be labelled more clearly.

None of this makes the method wrong. The architecture is plausible, the loss is novel, and the reported numbers are internally consistent in the ablations. But the abstract's 'state-of-the-art AP' and '4-8x speedup' are not supported by the evidence as presented. The authors need to either provide a same-backbone baseline, run everything on one GPU, or soften the claims.

Bottom line: if you're working in this subfield, cite it for the ranking loss and the feature-reuse direction, but don't trust the speedup numbers. It deserves a serious referee—the core idea is sound and the evaluation issues are fixable. I'd recommend sending it to review with a request for fair comparisons and variance estimates.

Referee Report

3 major / 3 minor

Summary. The paper proposes RARE, a lightweight traffic accident anticipation framework that reuses intermediate features from a single pre-trained YOLOv10 object detector. The pipeline extracts scene-level temporal context through a GRU over backbone features, computes object-centric embeddings with RoI Align and CBAM, fuses them with multi-head attention, and predicts accident risk from a short queue of fused features. The authors introduce an Attention Score Ranking Loss intended to force higher attention scores on accident-relevant objects, and an AdaLEA-based training objective. On the DAD and CCD benchmarks, they report state-of-the-art Average Precision (62.2% and 99.8%) at 73.3 FPS on an RTX 6000 Ada, claiming a 4-8x speedup over prior methods. Ablations show that removal of the ranking loss or of either feature stream reduces AP by roughly 6-8 points.

Significance. If substantiated, the result is practically valuable: a real-time accident anticipation system that avoids optical flow and additional feature extractors could substantially lower deployment cost, and the attention interpretability is a useful property for safety-critical applications. The paper is also clearly written and evaluates on standard benchmarks with standard metrics. I found no circularity in the evaluation: AP and mTTA are standard held-out metrics, and hyperparameters are selected on validation. However, the two central claims--the 4-8x speedup and the attribution of accuracy gains to RARE's design--are not currently supported by the evidence because the comparison mixes hardware, preprocessing, and detector backbones. In addition, the ranking loss as written in Eq. (5) is inverted and cannot behave as described. These issues are load-bearing and require a revision that includes controlled experiments and a corrected loss.

major comments (3)

[Section 4.1, Table 1] The claimed 4-8x speedup is not established because the FPS values for the baselines are taken from original papers running on different hardware and with different preprocessing, as acknowledged by the * footnote. The GSC number is adjusted by separately measured overhead, but this still does not place all methods on the same GPU, same input pipeline, or same detector. To support the central real-time claim, the authors should run at least the strongest baselines (e.g., GSC, DSTA, UString) on the same RTX 6000 Ada GPU with the same 640x640 preprocessing and, where possible, the same YOLOv10 backbone. Without such controlled measurements, the speed advantage could be attributable to hardware generation or to the highly optimized detector rather than to feature reuse.
[Section 3.3, Eq. (5)] The Attention Score Ranking Loss is written as L_R = min(0, max(S_na) + m - min(S_a)). This is the opposite of a hinge ranking loss. When a negative box violates the margin (max(S_na) + m > min(S_a)), the argument is positive and min(0, .) returns zero, so the loss is exactly zero at the violation; when the margin is satisfied, the loss is negative and the gradient encourages only further satisfaction. The standard formulation is L_R = max(0, max(S_na) + m - min(S_a)). Since Table 3 attributes a 5.9 AP improvement to this loss, the authors must correct the equation and verify that the implemented loss matches the intended one.
[Section 4.4, Table 1] The accuracy comparison does not control for the backbone. RARE uses YOLOv10, while the compared methods were published with different detection or feature backbones, and no baseline is re-implemented with the same YOLOv10 features. The reported AP advantage over the strongest baseline GSC is 1.8 points on DAD (62.2 vs. 60.4) and 0.5 points on CCD (99.8 vs. 99.3), so the gain could plausibly come from the detector rather than from the Attention Score Ranking Loss or feature reuse. At minimum, the authors should ablate the detector choice (e.g., replace YOLOv10 with the detector used by a baseline, or insert YOLOv10 features into a baseline pipeline) to support the attribution.

minor comments (3)

[Table 2 vs. Table 1] The DAD numbers for UString and DSTA differ between Table 1 and Table 2 (e.g., UString AP 53.7 vs. 58.6, mTTA 3.53 vs. 2.12). The text says these are "results reported in [7]" after further training, but the table caption does not explain the difference; please add a note so readers are not confused.
[Section 4.1] There is a missing space in "confidence above0.1"; please correct the typographical error.
[Section 3.2, Eq. (2)] Equation (2) passes both F_backbone and F_neck to a single RoIAlign call, but RoIAlign usually consumes one feature map. Please clarify how the multi-scale features are combined (e.g., concatenation or separate RoIAligns before fusion).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: RARE's predictions are supervised by ground-truth labels and evaluated on held-out test sets; the speed comparison issue is a benchmarking confound, not a circular derivation.

full rationale

The paper's derivation chain is self-contained and does not reduce to its own inputs. The accident risk score y_t is produced by a classifier over a temporal queue of fused features (Eq. 7), and the training objective (Eq. 8) combines an external AdaLEA loss with the proposed Attention Score Ranking Loss. The ranking loss supervises object attention using ground-truth accident-related bounding boxes via IoU > 0.5, which is a legitimate form of supervised training, not a fitted parameter disguised as a prediction. All reported accuracy metrics (AP, mTTA) are computed on held-out test splits of DAD and CCD using standard evaluation protocols, so the central claims of accuracy and earliness are externally validated rather than defined by construction. The paper cites no self-authored prior work as load-bearing evidence; its components (YOLOv10, GRU, RoIAlign, CBAM, AdaLEA) are external and explicitly referenced. The only notable weakness is the comparison of FPS/latency numbers across different GPUs and preprocessing pipelines, which the paper itself flags with asterisks and separately measured overhead for GSC. This is a benchmarking fairness concern about hardware confounds, not a logical circularity where a prediction is equivalent to its input. Accordingly, no circular step meets the evidentiary standard required for a positive finding.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The framework reuses a pre-trained object detector's features, so its success is conditioned on the transferability of those features and the availability of accident-object annotations. No new physical entities or forces are introduced. The main free parameters are training hyperparameters and thresholds that affect the ranking loss and detection stage.

free parameters (5)

margin m = 0.1
Set empirically to control the margin in the attention ranking loss (Sec. 3.3; Sec. 4.1).
loss weighting gamma = 10
Balances AdaLEA loss and ranking loss in the final objective (Eq. 8; Sec. 4.1).
temporal queue size k = 10
Number of past fused features fed to the classifier (Sec. 3.4; Sec. 4.1).
positive/negative IoU threshold = 0.5
Defines accident-related boxes for the ranking loss supervision (Sec. 3.3).
detector confidence threshold = 0.1
Filters object detections before RoI alignment (Sec. 4.1).

assumptions (3)

domain assumption The pre-trained YOLO detector provides useful intermediate features for accident anticipation without fine-tuning.
The entire framework depends on reusing backbone and neck features from YOLO (Eq. 1-2); if these features are not informative for accident cues, the method fails. No validation of this transferability is provided beyond the final AP.
domain assumption Ground-truth accident-related bounding boxes are available in DAD and used for supervision of the ranking loss.
The ranking loss assigns positives by IoU with accident-related boxes (Sec. 3.3); this assumes the benchmark annotations are reliable and representative.
domain assumption The evaluation protocol (AP, mTTA) aligns with prior TAA benchmarks.
Results are compared against published numbers; any difference in evaluation code or threshold handling could affect comparability.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Real-time Traffic Accident Anticipation with Feature Reuse." pith.science (2026). https://pith.science/paper/BFCROFDZ

@misc{pith2026250517449,
  author       = {Pith},
  title        = {Pith review of: Real-time Traffic Accident Anticipation with Feature Reuse},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BFCROFDZ}},
  note         = {Machine review of arXiv:2505.17449}
}

read the original abstract

This paper addresses the problem of anticipating traffic accidents, which aims to forecast potential accidents before they happen. Real-time anticipation is crucial for safe autonomous driving, yet most methods rely on computationally heavy modules like optical flow and intermediate feature extractors, making real-world deployment challenging. In this paper, we thus introduce RARE (Real-time Accident anticipation with Reused Embeddings), a lightweight framework that capitalizes on intermediate features from a single pre-trained object detector. By eliminating additional feature-extraction pipelines, RARE significantly reduces latency. Furthermore, we introduce a novel Attention Score Ranking Loss, which prioritizes higher attention on accident-related objects over non-relevant ones. This loss enhances both accuracy and interpretability. RARE demonstrates a 4-8 times speedup over existing approaches on the DAD and CCD benchmarks, achieving a latency of 13.6ms per frame (73.3 FPS) on an RTX 6000. Moreover, despite its reduced complexity, it attains state-of-the-art Average Precision and reliably anticipates imminent collisions in real time. These results highlight RARE's potential for safety-critical applications where timely and explainable anticipation is essential.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 18 canonical work pages

[1]

Milestones in autonomous driv- ing and intelligent vehicles: Survey of surveys,

Long Chen et al., “Milestones in autonomous driv- ing and intelligent vehicles: Survey of surveys,” IEEE Transactions on Intelligent V ehicles, vol. 8, no. 2, pp. 1046–1056, 2022

work page 2022
[2]

World Health Organization, Global status report on road safety 2018, World Health Organization, 2019

work page 2018
[3]

A comprehensive review on limitations of autonomous driving and its im- pact on accidents and collisions,

Amit Chougule, Vinay Chamola, Aishwarya Sam, Fei Richard Yu, and Biplab Sikdar, “A comprehensive review on limitations of autonomous driving and its im- pact on accidents and collisions,” IEEE Open Journal of V ehicular Technology, 2023

work page 2023
[4]

Anticipating accidents in dashcam videos,

Fu-Hsiang Chan, Yu-Ting Chen, Yu Xiang, and Min Sun, “Anticipating accidents in dashcam videos,” in Computer Vision–ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part IV 13 . Springer, 2017, pp. 136–153

work page 2016
[5]

Uncertainty-based traffic accident anticipation with spatio-temporal rela- tional learning,

Wentao Bao, Qi Yu, and Yu Kong, “Uncertainty-based traffic accident anticipation with spatio-temporal rela- tional learning,” in Proceedings of the 28th ACM In- ternational Conference on Multimedia, 2020, pp. 2682– 2690

work page 2020
[6]

An attention-guided multistream feature fusion network for early localization of risky traffic agents in driving videos,

Muhammad Monjurul Karim, Zhaozheng Yin, and Ruwen Qin, “An attention-guided multistream feature fusion network for early localization of risky traffic agents in driving videos,” IEEE Transactions on Intelli- gent V ehicles, vol. 9, no. 1, pp. 1792–1803, 2023

work page 2023
[7]

However, this improvement comes at the cost of a significant mTTA drop, approximately 1.5 s

indicate that after further training, both DSTA [8] and UString [5] achieve higher AP, approaching 60 %. However, this improvement comes at the cost of a significant mTTA drop, approximately 1.5 s. In contrast, RARE not only sur- passes 60 % AP but also maintains a higher mTTA than these methods, demonstrating a superior balance between detection accuracy...

work page 2025
[8]

Gsc: A graph and spatio-temporal continuity based framework for ac- cident anticipation,

Tianhang Wang, Kai Chen, Guang Chen, Bin Li, Zhijun Li, Zhengfa Liu, and Changjun Jiang, “Gsc: A graph and spatio-temporal continuity based framework for ac- cident anticipation,” IEEE Transactions on Intelligent V ehicles, vol. 9, no. 1, pp. 2249–2261, 2023

work page 2023

Show all 22 references

[9]

A dynamic spatial-temporal attention network for early anticipation of traffic accidents,

Muhammad Monjurul Karim, Yu Li, Ruwen Qin, and Zhaozheng Yin, “A dynamic spatial-temporal attention network for early anticipation of traffic accidents,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 7, pp. 9590–9600, 2022

2022
[10]

Very deep convolutional networks for large-scale image recognition,

Karen Simonyan, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014

2014 arXiv
[11]

Faster r-cnn: Towards real-time object detection with region proposal networks,

Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” IEEE transactions on pattern analysis and machine intelligence , vol. 39, no. 6, pp. 1137–1149, 2016

2016
[12]

Anticipating traffic accidents with adaptive loss and large-scale incident db,

Tomoyuki Suzuki, Hirokatsu Kataoka, Yoshimitsu Aoki, and Yutaka Satoh, “Anticipating traffic accidents with adaptive loss and large-scale incident db,” in Pro- ceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3521–3529

2018
[13]

Unsupervised traffic accident de- tection in first-person videos,

Yu Yao, Mingze Xu, Yuchen Wang, David J Crandall, and Ella M Atkins, “Unsupervised traffic accident de- tection in first-person videos,” in IROS. IEEE, 2019

2019
[14]

Dota: Unsu- pervised detection of traffic anomaly in driving videos,

Yu Yao, Xizi Wang, Mingze Xu, Zelin Pu, Yuchen Wang, Ella Atkins, and David J Crandall, “Dota: Unsu- pervised detection of traffic anomaly in driving videos,” IEEE transactions on pattern analysis and machine in- telligence, vol. 45, no. 1, pp. 444–459, 2022

2022
[15]

Drive: Deep rein- forced accident anticipation with visual explanation,

Wentao Bao, Qi Yu, and Yu Kong, “Drive: Deep rein- forced accident anticipation with visual explanation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 7619–7628

2021
[16]

Toward explainable artificial intelligence for early an- ticipation of traffic accidents,

Muhammad Monjurul Karim, Yu Li, and Ruwen Qin, “Toward explainable artificial intelligence for early an- ticipation of traffic accidents,” Transportation research record, vol. 2676, no. 6, pp. 743–755, 2022

2022
[17]

Drama: Joint risk localiza- tion and captioning in driving,

Srikanth Malla, Chiho Choi, Isht Dwivedi, Joon Hee Choi, and Jiachen Li, “Drama: Joint risk localiza- tion and captioning in driving,” in Proceedings of the IEEE/CVF winter conference on applications of com- puter vision, 2023, pp. 1043–1052

2023
[18]

Abductive ego-view accident video understanding for safe driving perception,

Jianwu Fang, Lei-lei Li, Junfei Zhou, Junbin Xiao, Hongkai Yu, Chen Lv, Jianru Xue, and Tat-Seng Chua, “Abductive ego-view accident video understanding for safe driving perception,” in CVPR, 2024

2024
[19]

Yolov10: Real-time end-to-end object detection,

Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding, “Yolov10: Real-time end-to-end object detection,” arXiv preprint arXiv:2405.14458, 2024

2024 arXiv
[20]

Mask r-cnn,

Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Girshick, “Mask r-cnn,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2961–2969

2017
[21]

Cbam: Convolutional block attention module,

Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon, “Cbam: Convolutional block attention module,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 3–19

2018
[22]

Agent-centric risk as- sessment: Accident anticipation and risky region local- ization,

Kuo-Hao Zeng, Shih-Han Chou, Fu-Hsiang Chan, Juan Carlos Niebles, and Min Sun, “Agent-centric risk as- sessment: Accident anticipation and risky region local- ization,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2017, pp. 2222–2230

2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Milestones in autonomous driv- ing and intelligent vehicles: Survey of surveys,

Long Chen et al., “Milestones in autonomous driv- ing and intelligent vehicles: Survey of surveys,” IEEE Transactions on Intelligent V ehicles, vol. 8, no. 2, pp. 1046–1056, 2022

work page 2022

[2] [2]

World Health Organization, Global status report on road safety 2018, World Health Organization, 2019

work page 2018

[3] [3]

A comprehensive review on limitations of autonomous driving and its im- pact on accidents and collisions,

Amit Chougule, Vinay Chamola, Aishwarya Sam, Fei Richard Yu, and Biplab Sikdar, “A comprehensive review on limitations of autonomous driving and its im- pact on accidents and collisions,” IEEE Open Journal of V ehicular Technology, 2023

work page 2023

[4] [4]

Anticipating accidents in dashcam videos,

Fu-Hsiang Chan, Yu-Ting Chen, Yu Xiang, and Min Sun, “Anticipating accidents in dashcam videos,” in Computer Vision–ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part IV 13 . Springer, 2017, pp. 136–153

work page 2016

[5] [5]

Uncertainty-based traffic accident anticipation with spatio-temporal rela- tional learning,

Wentao Bao, Qi Yu, and Yu Kong, “Uncertainty-based traffic accident anticipation with spatio-temporal rela- tional learning,” in Proceedings of the 28th ACM In- ternational Conference on Multimedia, 2020, pp. 2682– 2690

work page 2020

[6] [6]

An attention-guided multistream feature fusion network for early localization of risky traffic agents in driving videos,

Muhammad Monjurul Karim, Zhaozheng Yin, and Ruwen Qin, “An attention-guided multistream feature fusion network for early localization of risky traffic agents in driving videos,” IEEE Transactions on Intelli- gent V ehicles, vol. 9, no. 1, pp. 1792–1803, 2023

work page 2023

[7] [7]

However, this improvement comes at the cost of a significant mTTA drop, approximately 1.5 s

indicate that after further training, both DSTA [8] and UString [5] achieve higher AP, approaching 60 %. However, this improvement comes at the cost of a significant mTTA drop, approximately 1.5 s. In contrast, RARE not only sur- passes 60 % AP but also maintains a higher mTTA than these methods, demonstrating a superior balance between detection accuracy...

work page 2025

[8] [8]

Gsc: A graph and spatio-temporal continuity based framework for ac- cident anticipation,

Tianhang Wang, Kai Chen, Guang Chen, Bin Li, Zhijun Li, Zhengfa Liu, and Changjun Jiang, “Gsc: A graph and spatio-temporal continuity based framework for ac- cident anticipation,” IEEE Transactions on Intelligent V ehicles, vol. 9, no. 1, pp. 2249–2261, 2023

work page 2023