Pith. sign in

REVIEW 3 major objections 4 minor 42 references

NSF-HRPT: Neural Semantic Field meets Hierarchical Risk Perception Tree for Safety-Critical Scenario Assessment

T0 review · 3 major / 4 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 A simulation-trained neural semantic field, queried through a hierarchical risk tree, yields per-agent collision-time distributions from monocular video.

desk verdict A systems paper worth refereeing, but the real-world TTC comparison is undermined by an mTTA gate that the baselines cannot satisfy. read the letter →

arxiv 2608.04776 v1 pith:F2NEE6NG submitted 2026-08-05 cs.AI

Yu Zhao , Jiangyu Pan , Tao Hu , Ming Yin , Fan Yang , Jiangfan Liu , Xiubo Liang This is my paper
classification cs.AI
keywords neuralsemanticfieldtime-to-collisiondistributionhierarchicalriskperceptiontreesafety-criticalscenarioassessmentsim-to-realtransfermonocularaccidentanticipationautonomousdriving
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that quantitative, per-agent collision risk can be read directly from monocular dashcam video by combining a learning-based scene representation with a structured spatial reasoning tree. Its central claim is that a Neural Semantic Field trained only in a high-fidelity traffic simulator can answer queries about scene semantics, future motion, and probabilistic Time-to-Collision for every visible road user, and that a Hierarchical Risk Perception Tree turns those query answers into real-time, spatially organized risk values. The payoff of the claim is practical: if true, an autonomous system can obtain physically grounded TTC estimates and risk localization without needing large collections of real near-miss or crash labels, which are rare and expensive. The paper reports state-of-the-art results on its synthetic benchmark and competitive, near-state-of-the-art results on the DAD and CCD real dashcam datasets.

What carries the argument

The load-bearing object is the Neural Semantic Field (NSF), a continuous query-based function that maps $(x, y, \Delta t, \tau)$ to semantic logits, trajectory offsets, and Gaussian TTC parameters; it is implemented by a spatio-temporal transformer encoder, an iterative neural attention field, and a multi-task decoder with conditional batch normalization. The companion mechanism is the Hierarchical Risk Perception Tree (HRPT), a parallel five-level quadtree that stores per-entity RiskValues at leaf nodes and aggregates the maximum risk upward while carrying 4-bit RiskMask and ActiveMask indicators. The identity that connects the two is the RiskValue formula $R = p/(\mu+\delta)$ with $p = 1/(1+\sigma^2)$, which converts the predicted TTC distribution into the scalar urgency score the tree propagates and thresholds.

What would settle it

Re-run DAD and CCD with each baseline under the paper's exact validity gate (a warning counts only when the predicted TTC is within 0.3 seconds of ground truth); if the baselines cannot produce a TTC at all, or if their mTTA values change materially, the claimed near-state-of-the-art real-world result is not an apples-to-apples comparison and the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim, stated on its own terms, is that a single query-conditioned Neural Semantic Field can jointly encode semantics, motion, and risk: given a query $(x, y, \Delta t, \tau)$ it outputs a semantic class, a trajectory offset, and a Gaussian Time-to-Collision distribution with mean $\mu$ and variance $\sigma^2$. Risk is then defined as $R = p/(\mu+\delta)$ with $p = 1/(1+\sigma^2)$, which combines confidence and urgency into one number per entity. The Hierarchical Risk Perception Tree organizes detected entities into a five-layer quadtree, propagating maximum risk upward and using 4-bit RiskMask and ActiveMask indicators to localize the dangerous quadrant; this hierarchy is what the authors say makes the per-agent reasoning parallel and fast enough for real-time use. The paper further claims that, without retraining, fusing frozen depth and semantic features from foundation models at the input or feature level substantially closes the sim-to-real gap on the DAD and CCD benchmarks.

Load-bearing premise

The reported real-world comparison assumes that the baseline methods' average warning times were computed under the same validity rule used here—predicted time-to-collision within 0.3 seconds of the true value—but several of those baselines do not output a time-to-collision, so their published numbers may follow a different rule, and the near-state-of-the-art mTTA claim depends on that assumption.

Editorial extensions

If this is right

  • If the central claim holds, accident anticipation systems can produce per-agent TTC distributions rather than only a scene-level accident score, which is the information a planner or driver warning system actually needs.
  • Because the field is trained exclusively on synthetic data, the same weights transfer to new real-world cameras by adding frozen depth and semantic features, avoiding collection of real crash labels.
  • The quadtree hierarchy gives a natural way to report not just that risk is high, but where the risk is: the RiskMask flags which spatial quadrants contain the dangerous entity.
  • On the authors' synthetic benchmark the method reports the best AP, mTTA, and AOLA among the compared methods, supporting the claim that the framework is state of the art in simulation.
  • The multi-task NSF outputs semantic occupancy, trajectory, and TTC from one set of weights, so the representation is reusable for scene understanding, motion prediction, and risk assessment simultaneously.

Reading between the lines

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

  • Editorial inference: the RiskValue formula $R = p/(\mu+\delta)$ is a heuristic fusion of confidence and urgency; since the model already outputs $\mu$ and $\sigma^2$, one could test alternative fusion rules, such as the lower quantile of the predicted TTC distribution, without retraining.
  • Editorial inference: the paper's validity gate for a valid warning requires predicted TTC to be within 0.3 seconds of ground truth, but several compared baselines do not output TTC; if their published numbers were computed under a different rule, the real-world ranking may not be directly comparable.
  • Editorial inference: because the NSF is query-based and continuous, a natural extension is to query risk at arbitrary resolution or at times beyond the 5-second prediction horizon tested here, which could let the same field serve planning loops at different lookahead scales.
  • Editorial inference: a closed-loop test in the simulator that uses the HRPT risk values as a controller's reward or trigger would show whether the predicted TTC distributions actually improve emergency braking or avoidance decisions, which the offline metrics do not directly measure.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes NSF-HRPT, a framework for monocular, simulation-trained risk assessment in safety-critical driving scenes. A Neural Semantic Field (NSF) is trained on CARLA data to answer BEV queries for semantic occupancy, trajectory offsets, and probabilistic Time-to-Collision (TTC); at inference a Hierarchical Risk Perception Tree (HRPT) aggregates per-entity risk values in a parallel quadtree, and a Sim2Real enhancement fuses frozen depth and semantic features without retraining. The authors evaluate on CARLA scenarios and on the DAD and CCD real-world dashcam benchmarks, reporting AP, mTTA, and AOLA against five risk-scoring baselines and the LLM-based method of Liao et al. The central claim is state-of-the-art synthetic performance and near-state-of-the-art real-world performance for both TTC estimation accuracy and risk localization precision.

Significance. If the claims held, the contribution would be practically significant: a real-time monocular system producing per-agent probabilistic TTC estimates with explicit uncertainty, trained only on synthetic data and transferring to real dashcam video through frozen feature fusion. The NSF query formulation and the hierarchical parallel risk aggregation are sensible design choices, and the paper is commendable for reporting AP and AOLA alongside timing metrics, since those two metrics are protocol-independent. However, no code or checkpoints are released, and the central TTC-accuracy claim rests on an mTTA comparison whose validity criterion cannot be satisfied by most baselines. The risk-localization part (AP and AOLA) is more defensible, but the headline TTC claim as stated is not supported by the reported experiments.

major comments (3)
  1. [Section 4.1.3, Tables 1 and 2] The mTTA comparison is invalid as stated. The protocol requires a valid warning only when the predicted RiskValue crosses threshold and |T_hat - T_GT| <= 0.3 s. DSA, adaLEA, Ustring, DSTA, and GG are accident-anticipation risk scorers that output probabilities or risk scores, not TTC estimates, so they cannot satisfy the second condition unless some additional mapping is introduced. The paper does not describe any such mapping, so the mTTA numbers for these baselines cannot be produced by the protocol defined in Section 4.1.3. This directly undermines the abstract's claim of 'near-state-of-the-art ... TTC estimation accuracy' on real-world datasets, which is supported only by the mTTA column of Table 2.
  2. [Section 1 and Section 4.1.3] Even for the proposed model, mTTA is not a direct measure of TTC estimation accuracy. mTTA is the average time between the first valid warning and the collision, which is a warning-timing metric, whereas the paper's own Section 1 motivates precise per-entity TTC estimation as distinct from scene-level mTTA. No direct TTC error metric (e.g., MAE, RMSE, or calibration error of the predicted TTC distribution) is reported for the model or for any baseline. The authors should report direct TTC errors against ground-truth TTC, or restrict the claims to warning-timing and risk-localization performance.
  3. [Section 3.1 Eq. (10) and Section 4.1.3] There is a partial self-reference in the mTTA protocol: the RiskValue R = p/(mu+delta) is a monotone function of the predicted TTC mean mu, and the validity gate requires |T_hat - T_GT| <= 0.3 s. Thus mTTA partly measures how consistent the model's TTC head is with its own risk ordering, not an independent judgment of risk quality. This is not fatal because AP and AOLA are computed under standard protocols and do not share the issue, but the paper should prominently separate the TTC-consistency claim from the independent risk-localization claim.
minor comments (4)
  1. [Table 2 caption] The caption contains a typo: 'Real-Wrold' should be 'Real-World'.
  2. [Figures 2 and 3] The text in the figures repeatedly uses 'RiskVlaue'; this should be 'RiskValue'.
  3. [Section 3.3] The paper cites [41] (Segment Anything) as the source of semantic segmentation features for the Sim2Real enhancement, but SAM is an instance segmentation model rather than a semantic segmentation model; the specific models used for the semantic and depth priors should be clarified.
  4. [Section 4.3, Figure 5] The ablation studies vary one epsilon value (0.5) but otherwise use hand-set thresholds (RTh, pTh, alpha_d, alpha_s, lambda weights) without sensitivity analysis or error bars; given that the reported performance differences in Figures 4 and 5 are small, the paper should report standard deviations or repeated-run statistics to show which differences are meaningful.

Circularity Check

1 steps flagged · score 4.0 of 10

The mTTA validity gate is defined from the model's own predicted TTC, making the headline TTC-accuracy comparison self-referential; AP and AOLA remain independent evidence.

  1. self definitional [Section 4.1.3 (Evaluation Metrics) and Table 2 (DAD/CCD comparison)]
    "we define tθ as a valid warning moment only when: (1) the RiskValue predicted for tθ +δt first crosses the threshold, and (2) the absolute error between the predicted TTC ( ˆTtθ+δt) and ground-truth TTC (TGT,t θ+δt) satisfies | ˆTtθ+δt −T GT,t θ+δt| ≤ϵ(with ϵ= 0.3 s based on autonomous driving requirements)."

    The abstract's 'TTC estimation accuracy' claim rests on Table 2's mTTA column, but this mTTA is not an independent anticipation-time measure: its validity gate is literally the model's own predicted TTC. RiskValue itself is a function of the same predicted TTC via Eq. (10), R = p·1/(µ+δ), so both gate conditions reduce to properties of the trained TTC head µ. Moreover, baselines DSA, adaLEA, Ustring, DSTA, and GG are risk-scoring models that do not output TTC, so under the paper's stated protocol their Table 2 mTTA values cannot be produced; the comparison therefore mixes a self-gated TTC-accuracy test with published risk-scoring numbers. No direct TTC error metric (e.g., MAE) is reported, so the claimed TTC accuracy is supported only by this self-referential gate.

full rationale

The core derivation is self-contained: the NSF is trained on CARLA with Eq. (6)-(9), HRPT performs query-based hierarchical risk aggregation, and the Sim2Real fusion uses frozen foundation-model features; real-world DAD/CCD evaluation is on held-out data, so the sim-to-real transfer claim is not circular by construction. The significant self-referential step is the mTTA protocol in Section 4.1.3, where a 'valid warning' requires the model's own predicted TTC to be within 0.3 s of ground truth, and RiskValue is a monotone transformation of that same predicted TTC. Because five of the six baselines in Table 2 do not output TTC, their reported mTTA values cannot follow the stated protocol, so the near-SOTA TTC-accuracy claim rests on an apples-to-oranges comparison. AP and AOLA provide independent, standard evaluation and keep the overall circularity partial rather than total. The self-citation to [19] for scenario selection is a methodological choice, not a load-bearing uniqueness argument.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central method relies on an implicit BEV query space trained only in CARLA, hand-set risk and fusion thresholds, and loss weights that are not reported. No new physical entities are proposed. The evaluation additionally relies on a custom mTTA validity gate tied to the model's own TTC output.

free parameters (8)
  • RTh = 0.3
    RiskValue threshold in the HRPT RiskMask (Eq. 11). Hand-set; no sensitivity analysis is reported, and it directly controls which warnings are counted.
  • pTh = 0.7
    Confidence probability threshold in RiskMask (Eq. 11). Hand-set; affects alert triggering.
  • alpha_d = 0.5
    Depth fusion weight in Eq. 14. Fixed without a described tuning or validation protocol.
  • alpha_s = 0.3
    Semantic fusion weight in Eq. 14. Fixed without a described tuning or validation protocol.
  • lambda_sem, lambda_traj, lambda_ttc = not reported
    Loss weights in Eq. 6 are said to balance the tasks but their values are not given; they must be tuned on validation data.
  • epsilon = 0.3 s
    TTC error tolerance in the mTTA validity criterion (Section 4.1.3). Hand-set and changes which warnings count.
  • delta = 0.1
    Numerical stability constant in RiskValue Eq. 10; minor but affects risk magnitude.
  • beta = 1.0
    Threshold in Smooth L1 trajectory offset loss (Eq. 8), predefined rather than tuned; minor.
assumptions (4)
  • domain assumption Monocular image features learned in CARLA encode enough 3D layout information for BEV coordinate queries to produce meaningful semantic and trajectory outputs.
    The NSF is queried at BEV coordinates (x,y) and trained only in simulation (Section 3.1, NSF for Inference). Real-world validity of this implicit geometry is assumed.
  • domain assumption Pretrained depth (DPT) and segmentation (SAM) features bridge the appearance gap so the frozen CARLA-trained NSF remains valid on DAD/CCD.
    Section 3.3 introduces early and mid fusion but provides no analysis showing the fused features align the BEV query space across domains.
  • domain assumption TTC ground truth computed in CARLA is a valid proxy for real-world collision urgency.
    Training and the mTTA validity gate (Eq. 9, Section 4.1.3) rely on simulated TTC labels; real datasets lack TTC ground truth in the same form.
  • ad hoc to paper RiskValue R = p / (mu + delta) is a faithful ordering of collision urgency.
    Eq. 10 defines risk as inverse predicted TTC scaled by a confidence term; this is a modeling choice with no external validation against collision outcomes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NSF-HRPT: Neural Semantic Field meets Hierarchical Risk Perception Tree for Safety-Critical Scenario Assessment." pith.science (2026). https://pith.science/paper/F2NEE6NG

@misc{pith2026260804776,
  author       = {Pith},
  title        = {Pith review of: NSF-HRPT: Neural Semantic Field meets Hierarchical Risk Perception Tree for Safety-Critical Scenario Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F2NEE6NG}},
  note         = {Machine review of arXiv:2608.04776}
}
read the original abstract

The ability to accurately assess and anticipate risks in safety-critical scenarios is crucial for autonomous driving systems. While existing research has made progress in collision prediction, accurately quantifying risk levels from monocular vision inputs remains challenging due to the complex dynamics of multi-agent interactions and the inherent uncertainty in real-world environments. To address these challenges, we present NSF-HRPT, a novel framework that combines learning-based perception with structured reasoning for quantitative risk assessment. Our approach features a Neural Semantic Field (NSF) that learns to model scene semantics, trajectory predictions, and probabilistic Time-to-Collision (TTC) distributions from simulation data. During inference, the pre-trained NSF serves as a prior for our Hierarchical Risk Perception Tree (HRPT), which enables efficient parallel computation and spatial reasoning about multi-agent risks. Additionally, we introduce a Sim2Real enhancement strategy that improves real-world applicability without retraining by incorporating priors from foundation models. Extensive evaluations demonstrate that our framework achieves state-of-the-art performance on synthetic benchmarks and delivers competitive, near-state-of-the-art results on real-world datasets for both TTC estimation accuracy and risk localization precision. The proposed method provides an effective solution for real-time risk awareness from monocular camera inputs.

Figures

Figures reproduced from arXiv: 2608.04776 by the authors.

Figure 1
Figure 1. Neural Semantic Field (NSF) Architecture. The NSF model comprises three core modules: a Spatio￾Temporal Encoder for feature extraction, a Neural Attention Field for iterative query-conditioned context refinement, and a Multi-task Decoder that outputs semantic occupancy, trajectory offsets, and probabilistic TTC distributions. The illustration uses T = 3 input frames and P = 4 spatial patches for demonstration. 3.1 N… view at source ↗
Figure 2
Figure 2. The Structure of HRPT. The quadtree organizes spatial risk information across five hierarchical layers (Root, two Internal, LeafFather, Leaf). The BEV view illustrates the spatial partitioning, while the tree diagram details the data (RiskValue, RiskMask, ActiveMask, EntityList, EntityID) stored at each node type, facilitating efficient parallel risk aggregation. • RiskValue ∈ R +: Risk value computed from NSF queri… view at source ↗
Figure 3
Figure 3. The Inference Process of HRPT. The diagram contrasts the tree’s state during initial safe conditions versus upon detecting a high-risk entity. The BEV views are for conceptual illustration; the framework operates on a logical BEV representation from the NSF without requiring explicit rendering. NOTE: The BEV visualization is a schematic for clarity; the actual algorithm uses an implicit, query-based spatial represen… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Component Analysis on DAD Dataset Using AP. "- Red/Green Lights": removing the ’red light’ and ’green [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Component Analysis on CARLA Using mTTA and AOLA. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages

  1. [1]

    A survey on safety-critical driving scenario generation—a methodological perspective.IEEE Transactions on Intelligent Transportation Systems, 2023

    Wenhao Ding, Chejian Xu, Mansur Arief, Haohong Lin, Bo Li, and Ding Zhao. A survey on safety-critical driving scenario generation—a methodological perspective.IEEE Transactions on Intelligent Transportation Systems, 2023

    work page 2023

  2. [2]

    Corner cases for visual perception in automated driving: Some guidance on detection approaches.arXiv preprint arXiv:2102.05897, 2021

    Jasmin Breitenstein, Jan-Aike Termöhlen, Daniel Lipinski, and Tim Fingscheidt. Corner cases for visual perception in automated driving: Some guidance on detection approaches.arXiv preprint arXiv:2102.05897, 2021

    arXiv 2021

  3. [3]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2020

    work page 2020

  4. [4]

    Womd-lidar: Raw sensor dataset benchmark for motion forecasting

    Kan Chen, Runzhou Ge, Hang Qiu, Rami Ai-Rfou, Charles Qi, Xuanyu Zhou, Zoey Yang, Scott Ettinger, Pei Sun, Zhaoqi Leng, et al. Womd-lidar: Raw sensor dataset benchmark for motion forecasting. InProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2024

    work page 2024

  5. [5]

    Nuscenes: A multimodal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. Nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2020

    work page 2020

  6. [6]

    Anticipating accidents in dashcam videos

    Fu-Hsiang Chan, Yu-Ting Chen, Yu Xiang, and Min Sun. Anticipating accidents in dashcam videos. InAsian Conference on Computer Vision, 2016

    work page 2016

  7. [7]

    Uncertainty-based traffic accident anticipation with spatio-temporal relational learning

    Wentao Bao, Qi Yu, and Yu Kong. Uncertainty-based traffic accident anticipation with spatio-temporal relational learning. InProceedings of the 28th ACM International Conference on Multimedia, 2020

    work page 2020

  8. [8]

    Safety-critical scenario generation via reinforcement learning based editing

    Haolan Liu, Liangjun Zhang, Siva Kumar Sastry Hari, and Jishen Zhao. Safety-critical scenario generation via reinforcement learning based editing. InIEEE International Conference on Robotics and Automation (ICRA), 2024

    work page 2024

Show all 42 references
  1. [9]

    Diffscene: Diffusion-based safety-critical scenario generation for autonomous vehicles

    Chejian Xu, Aleksandr Petiushko, Ding Zhao, and Bo Li. Diffscene: Diffusion-based safety-critical scenario generation for autonomous vehicles. InProceedings of the AAAI Conference on Artificial Intelligence, 2025

    2025

  2. [10]

    Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles

    Jiawei Zhang, Chejian Xu, and Bo Li. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. InProceedings of the Computer Vision and Pattern Recognition Conference, 2024

    2024

  3. [11]

    A dynamic spatial-temporal attention network for early anticipation of traffic accidents.IEEE Transactions on Intelligent Transportation Systems, 2022

    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, 2022

    2022

  4. [12]

    World model- based end-to-end scene generation for accident anticipation in autonomous driving.Communications Engineering, 2025

    Yanchen Guan, Haicheng Liao, Chengyue Wang, Xingcheng Liu, Jiaxun Zhang, and Zhenning Li. World model- based end-to-end scene generation for accident anticipation in autonomous driving.Communications Engineering, 2025. 11 NSF-HRPTPREPRINT

    2025

  5. [13]

    CARLA: An open urban driving simulator

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. InProceedings of the 1st Annual Conference on Robot Learning (CoRL), 2017

    2017

  6. [14]

    Diffroad: Realistic and diverse road scenario generation for autonomous vehicle testing.arXiv preprint arXiv:2411.09451, 2024

    Junjie Zhou, Lin Wang, Qiang Meng, and Xiaofan Wang. Diffroad: Realistic and diverse road scenario generation for autonomous vehicle testing.arXiv preprint arXiv:2411.09451, 2024

    2024 arXiv

  7. [15]

    Trafficgen: Learning to generate diverse and realistic traffic scenarios

    Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and realistic traffic scenarios. In2023 IEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  8. [16]

    Authsim: Towards authentic and effective safety-critical scenario generation for autonomous driving tests

    Yukuan Yang, Xucheng Lu, Zhili Zhang, Zepeng Wu, Guoqi Li, Lingzhong Meng, Zhiming Ding, and Yunzhi Xue. Authsim: Towards authentic and effective safety-critical scenario generation for autonomous driving tests. arXiv preprint arXiv:2502.21100, 2025

    2025 arXiv

  9. [17]

    Safety2drive: Safety-critical scenario benchmark for the evaluation of autonomous driving.arXiv preprint arXiv:2505.13872, 2025

    Jingzheng Li, Tiancheng Wang, Xingyu Peng, Jiacheng Chen, Zhijun Chen, Bing Li, and Xianglong Liu. Safety2drive: Safety-critical scenario benchmark for the evaluation of autonomous driving.arXiv preprint arXiv:2505.13872, 2025

    2025 arXiv

  10. [18]

    Seeking to collide: Online safety-critical scenario generation for autonomous driving with retrieval augmented large language models.arXiv preprint arXiv:2505.00972, 2025

    Yuewen Mei, Tong Nie, Jian Sun, and Ye Tian. Seeking to collide: Online safety-critical scenario generation for autonomous driving with retrieval augmented large language models.arXiv preprint arXiv:2505.00972, 2025

    2025 arXiv

  11. [19]

    Adversarial generation and collaborative evolution of safety-critical scenarios for autonomous vehicles.arXiv preprint arXiv:2508.14527, 2025

    Jiangfan Liu, Yongkang Guo, Fangzhi Zhong, Tianyuan Zhang, Zonglei Jing, Siyuan Liang, Jiakai Wang, Mingchuan Zhang, Aishan Liu, and Xianglong Liu. Adversarial generation and collaborative evolution of safety-critical scenarios for autonomous vehicles.arXiv preprint arXiv:2508...

    2025 arXiv

  12. [20]

    Egocentric vision-based future vehicle localization for intelligent driving assistance systems

    Yu Yao, Mingze Xu, Chiho Choi, David J Crandall, Ella M Atkins, and Behzad Dariush. Egocentric vision-based future vehicle localization for intelligent driving assistance systems. InIEEE International Conference on Robotics and Automation (ICRA), 2019

    2019

  13. [21]

    Unsupervised traffic accident detection in first-person videos

    Yu Yao, Mingze Xu, Yuchen Wang, David J Crandall, and Ella M Atkins. Unsupervised traffic accident detection in first-person videos. InIEEE/RSJ International Conference on Intelligent Robots and Systems, 2019

    2019

  14. [22]

    Anovox: A benchmark for multimodal anomaly detection in autonomous driving

    Daniel Bogdoll, Iramm Hamdard, Lukas Namgyu Rößler, Felix Geisler, Muhammed Bayram, Felix Wang, Jan Imhof, Miguel De Campos, Anushervon Tabarov, Yitian Yang, et al. Anovox: A benchmark for multimodal anomaly detection in autonomous driving. InEuropean Conference on Computer Vi...

    2024

  15. [23]

    Spotting the unexpected (stu): A 3d lidar dataset for anomaly segmentation in autonomous driving

    Alexey Nekrasov, Malcolm Burdorf, Stewart Worrall, Bastian Leibe, and Julie Stephany Berrio Perez. Spotting the unexpected (stu): A 3d lidar dataset for anomaly segmentation in autonomous driving. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025

    2025

  16. [24]

    Umad: Unsupervised mask-level anomaly detection for autonomous driving.arXiv preprint arXiv:2406.06370, 2024

    Daniel Bogdoll, Noël Ollick, Tim Joseph, Svetlana Pavlitska, and J Marius Zöllner. Umad: Unsupervised mask-level anomaly detection for autonomous driving.arXiv preprint arXiv:2406.06370, 2024

    2024 arXiv

  17. [25]

    Jisam: Alleviate labeling burden and corner case problems in autonomous driving via minimal real-world data

    Runjian Chen, Wenqi Shao, Bo Zhang, Shaoshuai Shi, Li Jiang, and Ping Luo. Jisam: Alleviate labeling burden and corner case problems in autonomous driving via minimal real-world data. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025

    2025

  18. [26]

    Graph(graph): A nested graph-based framework for early accident anticipation

    Nupur Thakur, PrasanthSai Gouripeddi, and Baoxin Li. Graph(graph): A nested graph-based framework for early accident anticipation. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024

    2024

  19. [27]

    Crash: Crash recognition and anticipation system harnessing with context-aware and temporal focus attentions

    Haicheng Liao, Haoyu Sun, Huanming Shen, Chengyue Wang, Chunlin Tian, KaHou Tam, Li Li, Chengzhong Xu, and Zhenning Li. Crash: Crash recognition and anticipation system harnessing with context-aware and temporal focus attentions. InProceedings of the 32nd ACM International Con...

    2024

  20. [28]

    Latte: A real-time lightweight attention-based traffic accident anticipation engine.Information Fusion, 2025

    Jiaxun Zhang, Yanchen Guan, Chengyue Wang, Haicheng Liao, Guohui Zhang, and Zhenning Li. Latte: A real-time lightweight attention-based traffic accident anticipation engine.Information Fusion, 2025

    2025

  21. [29]

    Singulartrajectory: Universal trajectory predictor using diffusion model

    Inhwan Bae, Young-Jae Park, and Hae-Gon Jeon. Singulartrajectory: Universal trajectory predictor using diffusion model. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2024

    2024

  22. [30]

    Egonav: Egocentric scene-aware human trajectory prediction.arXiv preprint arXiv:2403.19026, 2024

    Weizhuo Wang, C Karen Liu, and Monroe Kennedy III. Egonav: Egocentric scene-aware human trajectory prediction.arXiv preprint arXiv:2403.19026, 2024

    2024 arXiv

  23. [31]

    A novel benchmarking paradigm and a scale-and motion-aware model for egocentric pedestrian trajectory prediction

    Amir Rasouli. A novel benchmarking paradigm and a scale-and motion-aware model for egocentric pedestrian trajectory prediction. InIEEE International Conference on Robotics and Automation (ICRA), 2024

    2024

  24. [32]

    Drivemrp: Enhancing vision-language models with synthetic motion data for motion risk prediction

    Zhiyi Hou, Enhui Ma, Fang Li, Zhiyi Lai, Kalok Ho, Zhanqian Wu, Lijun Zhou, Long Chen, Chitian Sun, Haiyang Sun, et al. Drivemrp: Enhancing vision-language models with synthetic motion data for motion risk prediction. arXiv preprint arXiv:2507.02948, 2025

    2025 arXiv

  25. [33]

    When, where, and what? a benchmark for accident anticipation and localization with large language models

    Haicheng Liao, Yongkang Li, Chengyue Wang, Yanchen Guan, Kahou Tam, Chunlin Tian, Li Li, Chengzhong Xu, and Zhenning Li. When, where, and what? a benchmark for accident anticipation and localization with large language models. InProceedings of the 32nd ACM International Confer...

    2024

  26. [34]

    Neat: Neural attention fields for end-to-end autonomous driving

    Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. InInternational Conference on Computer Vision (ICCV), 2021

    2021

  27. [35]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2016

    2016

  28. [36]

    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. InProceedings of the Neural Information Processing Systems (NeurIPS), 2017

    2017

  29. [37]

    Safebench: A benchmarking platform for safety evaluation of autonomous vehicles

    Chejian Xu, Wenhao Ding, Weijie Lyu, Zuxin Liu, Shuai Wang, Yihan He, Hanjiang Hu, Ding Zhao, and Bo Li. Safebench: A benchmarking platform for safety evaluation of autonomous vehicles. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  30. [38]

    Najm, John D

    Wassim G. Najm, John D. Smith, and Mikio Yanagisawa. Pre-crash scenario typology for crash avoidance research. Technical report, John A. V olpe National Transportation Systems Center (U.S.) (V olpe), 2007. United States. National Highway Traffic Safety Administration

    2007

  31. [39]

    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. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2020

    2020

  32. [40]

    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. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2018

    2018

  33. [41]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. InIEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  34. [42]

    Vision transformers for dense prediction

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 13

    2021

Pith tools

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