Beyond Fixed Thresholds and Domain-Specific Benchmarks for Explainable Multi-Task Classification in Autonomous Vehicles

Maryam Sadat Hosseini Azad; Shahriar Baradaran Shokouhi

arxiv: 2605.04299 · v1 · submitted 2026-05-05 · 💻 cs.CV · cs.RO

Beyond Fixed Thresholds and Domain-Specific Benchmarks for Explainable Multi-Task Classification in Autonomous Vehicles

Maryam Sadat Hosseini Azad , Shahriar Baradaran Shokouhi This is my paper

Pith reviewed 2026-05-08 17:09 UTC · model grok-4.3

classification 💻 cs.CV cs.RO

keywords explainable AIautonomous vehiclesmulti-task classificationthreshold sensitivity analysisdriving decision datasetscene understandingcross-cultural evaluation

0 comments

The pith

Adaptive threshold selection raises F1 scores in multi-task explainable classification for autonomous driving.

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

The paper shows that fixed confidence thresholds perform poorly when a single model must simultaneously classify driving behaviors and generate their explanations. Through sensitivity analysis the authors identify task-specific thresholds that improve F1 scores across predictions. They also release the IUST-XAI-AD dataset of 958 human-annotated images that captures driving decisions and reasoning in varied contexts. This combination addresses the lack of transparent, culturally aware evaluation tools for black-box deep learning systems in vehicles. If the results hold, both threshold tuning and domain-specific benchmarks become required steps for building trustworthy autonomous perception.

Core claim

The authors establish that traditional fixed thresholds are suboptimal for multi-task scenarios in explainable autonomous driving perception; an adaptive threshold selection methodology derived from confidence sensitivity analysis improves F1-scores, while the new IUST-XAI-AD dataset of 958 annotated images reveals cross-cultural driving behavior patterns and provides a more challenging benchmark than prior resources.

What carries the argument

The adaptive threshold selection methodology, which evaluates multiple confidence values to set task-specific decision boundaries that maximize F1 performance in simultaneous driving behavior and explanation predictions.

If this is right

Multi-task models for scene understanding can be tuned to deliver higher accuracy on both behavior detection and explanation generation.
Evaluation protocols for autonomous driving systems must incorporate per-task threshold optimization instead of assuming one universal value.
The IUST-XAI-AD dataset enables measurement of cultural variation in driving decisions and reasoning that current benchmarks miss.
Deployable explainable systems for global use require both methodological advances in threshold handling and expanded domain-specific test data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The method may lower false-positive rates in safety-critical decisions by allowing each task to operate at its own optimal operating point.
Wider adoption of the new dataset could expose systematic biases in models trained only on data from limited geographic or cultural sources.
Real-time implementation would need to check whether the added sensitivity analysis step introduces unacceptable latency in onboard inference.

Load-bearing premise

The sensitivity analysis performed on the chosen models and tasks will generalize to other architectures and real-world driving distributions without further validation.

What would settle it

Retraining the multi-task model on a different architecture or a larger real-world driving dataset and observing that no adaptive threshold set outperforms the best fixed threshold would falsify the central performance claim.

Figures

Figures reproduced from arXiv: 2605.04299 by Maryam Sadat Hosseini Azad, Shahriar Baradaran Shokouhi.

**Figure 1.** Figure 1: The overview of the studied problem. S refers to sigmoid, O refers to output, T refers to threshold, and P refers to prediction. The workflow processes a dash-cam image through an interpretable deep learning model. The model outputs a sigmoid probability score S(O) that is compared to a threshold T, producing a binary classification prediction P. that not only perform well statistically but also operate sa… view at source ↗

**Figure 2.** Figure 2: Several frames from the IUST-XAI-AD dataset captured at different locations and times of day, demonstrating the high complexity and diversity of the dataset. about different object types in autonomous driving scenarios. The complexity score C is calculated as: 𝐶 = 1.5 × 𝐷𝑝 + 1.3 × 𝐷𝑟 + 1.0 × 𝐷𝑣 (3) Where 𝐷𝑝 , 𝐷𝑟 , and 𝐷𝑣 represent the density (objects per image) of pedestrians, riders, and vehicles, respe… view at source ↗

**Figure 3.** Figure 3: F1 score performance comparison between action (blue lines) and reason (red lines) classification tasks across different confidence thresholds. • Increasing Confidence Threshold: (1) Precision increases: When raising the threshold, the model only assigns positive labels to predictions with high confidence scores. This approach reduces the number of false positives, as the model becomes more selective in … view at source ↗

**Figure 6.** Figure 6: Detailed analysis of object type distributions across the three datasets, displaying Pedestrian Density, Rider Density, Vehicle Density per image, and a comparative overview of all object densities. IUST-XAI-AD shows the highest vehicle density (1.658) and rider density (0.164), while maintaining competitive pedestrian density (0.089) compared to other datasets. The analysis reveals significant cultural a… view at source ↗

**Figure 8.** Figure 8: shows two-dimensional t-SNE embedding of features colored by driving action classes: "Move forward" (pink, n=2475) and "Stop/Slow down" (teal, n=2095). The visualization shows clear clustering patterns with some overlap between classes, indicating that the model learns discriminative features while capturing the continuous nature of driving decisions. The t-SNE embedding for the reasons related to "stop/… view at source ↗

read the original abstract

Scene understanding is a vital part of autonomous driving systems, which requires the use of deep learning models. Deep learning methods are intrinsically black box models, which lack transparency and safety in autonomous driving. To make these systems transparent, multi-task visual understanding has become crucial for explainable autonomous driving perception systems, where simultaneous prediction of multiple driving behaviors and their underlying explanations is essential for safe navigation and human trust in autonomous vehicles. In order to design an accurate and cross-cultural explainable autonomous driving system, we introduce a comprehensive confidence threshold sensitivity analysis that evaluates various threshold values to identify optimal decision boundaries for different tasks. Our analysis demonstrates that traditional fixed threshold approaches are suboptimal for multi-task scenarios. Through extensive evaluation, we demonstrate that our adaptive threshold selection methodology improves F1-scores across different tasks. In addition, we introduce IUST-XAI-AD, a novel dataset consisting of 958 images with human annotations for driving decisions and corresponding reasoning. This dataset addresses the critical gap in domain-specific evaluation benchmarks for distinct driving contexts and provides a more challenging test environment compared to existing datasets. Experimental results demonstrate that confidence threshold sensitivity analysis can significantly improve model performance, while the introduction of the IUST-XAI-AD dataset reveals important insights about cross-cultural driving behavior patterns. The combined contributions of this work provide both methodological advances and practical evaluation tools that can accelerate the development of more reliable, explainable, and culturally-adaptive autonomous driving systems for global deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper adds a modest new 958-image dataset for explainable AV perception and claims adaptive thresholds lift F1 scores in multi-task settings, but the evidence for the gains is not visible in the abstract and the dataset is too small for strong claims.

read the letter

The core new thing here is the IUST-XAI-AD dataset of 958 images with human annotations on driving decisions and reasoning. That fills a narrow gap for domain-specific benchmarks in autonomous vehicle explainability, especially if the annotations capture cross-cultural patterns that existing sets miss. The methodological piece is a sensitivity analysis that picks task-specific thresholds instead of using one fixed value, with the claim that this raises F1 scores across tasks. If the full paper shows clean comparisons and the adaptation rule is simple and reproducible, that is a practical tweak worth noting for people building multi-task perception stacks in cars. The focus on transparency for safety and trust is reasonable given the application. The dataset release itself is the part that could see some use in follow-on work. The main weakness is that the abstract gives no numbers, no deltas, no variance, and no description of whether the threshold choice is static per task, per-sample, or learned. With only 958 images the risk of post-hoc tuning on the test set is real, and without cross-validation or statistical tests the reported improvement could be noise. The stress-test note is fair on this point. The paper does not appear to overclaim beyond what the experiments (whatever they are) support, but the lack of visible controls makes it hard to judge how general the result is. This is the kind of incremental CV-for-AV paper that a reading group might skim for the dataset details rather than the method. It is not strong enough on its own for a top-tier venue without tighter experiments, but the dataset and the practical motivation mean it could clear a serious referee process if the authors add the missing quantitative tables and validation steps. I would send it out for review rather than desk-reject, mainly to get feedback on the dataset utility.

Referee Report

3 major / 0 minor

Summary. The manuscript presents a confidence threshold sensitivity analysis for multi-task classification in explainable autonomous driving systems. It argues that fixed thresholds are suboptimal and proposes an adaptive threshold selection approach that purportedly improves F1-scores. The authors also introduce the IUST-XAI-AD dataset comprising 958 images annotated with driving decisions and explanations to fill gaps in domain-specific benchmarks for cross-cultural evaluation.

Significance. If substantiated with detailed experiments, the adaptive threshold method could offer a practical way to optimize multi-task performance in safety-critical AV applications, potentially increasing trust through better explainability. The new dataset may provide a challenging testbed for future XAI methods in autonomous driving, particularly for studying cultural variations in driving behavior. These contributions, if validated, address important practical challenges in deploying reliable perception systems globally.

major comments (3)

[Abstract] Abstract: The claim that 'our adaptive threshold selection methodology improves F1-scores across different tasks' is not accompanied by any quantitative results, baseline comparisons, variance measures, or statistical significance tests. This makes the central empirical claim difficult to evaluate.
[Abstract] Abstract: With a dataset of only 958 images, the sensitivity analysis for 'optimal decision boundaries' risks overfitting if thresholds are tuned on the evaluation set without cross-validation or held-out data; no details on data splitting or multiple runs are provided.
[Abstract] Abstract: The exact nature of the 'adaptive threshold selection' is not described: it is unclear if it is a per-task static optimization, a learned parameter, or instance-dependent, which is load-bearing for reproducibility and generalizability claims.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough review and constructive comments on our manuscript. We address each of the major comments point by point below, indicating the changes we plan to make in the revised version.

read point-by-point responses

Referee: [Abstract] Abstract: The claim that 'our adaptive threshold selection methodology improves F1-scores across different tasks' is not accompanied by any quantitative results, baseline comparisons, variance measures, or statistical significance tests. This makes the central empirical claim difficult to evaluate.

Authors: We agree that the abstract does not provide the quantitative details supporting this claim. While the full manuscript contains experimental results including F1-score comparisons in tables and figures, we will revise the abstract to include specific quantitative improvements, such as the percentage increase in F1-scores for each task, and mention that results are averaged over multiple runs with reported standard deviations. Additionally, we will include a note on statistical significance testing in the revised abstract. revision: yes
Referee: [Abstract] Abstract: With a dataset of only 958 images, the sensitivity analysis for 'optimal decision boundaries' risks overfitting if thresholds are tuned on the evaluation set without cross-validation or held-out data; no details on data splitting or multiple runs are provided.

Authors: This is a valid concern given the dataset size. The manuscript describes the IUST-XAI-AD dataset but omits detailed splitting information in the abstract. We will add explicit details on the data partitioning strategy, confirming the use of a held-out test set and cross-validation on the training/validation portions for threshold tuning. We will also report results from multiple independent runs to provide variance measures and mitigate overfitting risks. revision: yes
Referee: [Abstract] Abstract: The exact nature of the 'adaptive threshold selection' is not described: it is unclear if it is a per-task static optimization, a learned parameter, or instance-dependent, which is load-bearing for reproducibility and generalizability claims.

Authors: We appreciate this observation as it highlights a lack of clarity in the abstract. Our adaptive threshold selection is implemented as a per-task static optimization: for each classification task, we conduct a sensitivity analysis by evaluating a range of threshold values on a validation set to select the one that maximizes the F1-score. This is not instance-dependent nor a learned parameter during training. We will update the manuscript with a clear description of this methodology, including an algorithm box or pseudocode, to enhance reproducibility. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents an empirical methodology for adaptive threshold selection via sensitivity analysis on a newly introduced 958-image dataset, asserting F1-score improvements over fixed thresholds. No equations, derivations, or load-bearing self-citations appear in the provided text. Claims rest on experimental comparisons against the introduced benchmark rather than any reduction of outputs to fitted inputs or prior author results by construction. This is a standard empirical contribution whose validity can be assessed externally via the dataset and methods.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical derivations, free parameters, or new postulated entities are described in the abstract; the work appears entirely empirical.

pith-pipeline@v0.9.0 · 5569 in / 1094 out tokens · 30003 ms · 2026-05-08T17:09:34.679751+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

44 extracted references · 5 canonical work pages · 1 internal anchor

[1]

J. Dong, S. Chen, S. Zong, T. Chen, S. Labi, Image transformer for explainable autonomous driving system, in: 2021 IEEE international intelligenttransportationsystemsconference(ITSC),IEEE,2021,pp. 2732–2737

2021
[2]

H. Sun, M. Li, Z. Cui, Y. Huang, H. Chen, Semantic shapley-based counterfactual explanations for end-to-end autonomous driving, En- gineering Applications of Artificial Intelligence 159 (2025) 111638

2025
[3]

J.Xie,Y.Zhang,Y.Qin,B.Wang,S.Dong,K.Li,Y.Xia, Ishuman- like decision making explainable? towards an explainable artificial intelligence for autonomous vehicles, Transportation Research In- terdisciplinary Perspectives 29 (2025) 101278

2025
[4]

Albrecht, Ahuman-centricmethodforgeneratingcausalexplanations in natural language for autonomous vehicle motion planning, arXiv preprint arXiv:2206.08783 (2022)

B.Gyevnar,M.Tamborski,C.Wang,C.G.Lucas,S.B.Cohen,S.V. Albrecht, Ahuman-centricmethodforgeneratingcausalexplanations in natural language for autonomous vehicle motion planning, arXiv preprint arXiv:2206.08783 (2022)

work page arXiv 2022
[5]

M. S. Hosseini Azad, S. Baradaran Shokouhi, A. A. Hamidi Imani, S. Atakishiyev, R. Goebel, An end-to-end decision-aware multi-scale attention-based model for explainable autonomous driving, 2026

2026
[6]

C.Cao,X.Chen,J.Wang,Q.Song,R.Tan,Y.-H.Li,Sgdcl:Semantic- guided dynamic correlation learning for explainable autonomous driving, in: 33rd International Joint Conference on Artificial Intel- ligence (IJCAI 2024), International Joint Conferences on Artificial Intelligence, 2024, pp. 596–604

2024
[7]

S. Meng, Y. Wang, Y. Cui, L.-P. Chau, Foundation model-assisted interpretable vehicle behavior decision making, Knowledge-Based Systems (2025) 113868

2025
[8]

Z.Fu,K.Jiang,Y.Xu,Y.Wang,T.Wen,H.Gao,Z.Zhong,D.Yang, Top-down attention-based mechanisms for interpretable autonomous driving, IEEE Transactions on Intelligent Transportation Systems (2024)

2024
[9]

D. Wang, C. Devin, Q.-Z. Cai, F. Yu, T. Darrell, Deep object-centric policies for autonomous driving, in: 2019 International Conference on Robotics and Automation (ICRA), IEEE, 2019, pp. 8853–8859

2019
[10]

T. Jing, H. Xia, R. Tian, H. Ding, X. Luo, J. Domeyer, R. Sherony, Z. Ding, Inaction: Interpretable action decision making for au- tonomous driving, in: European Conference on Computer Vision, Springer, 2022, pp. 370–387

2022
[11]

F. Yu, H. Chen, X. Wang, W. Xian, Y. Chen, F. Liu, V. Madhavan, T. Darrell, Bdd100k: A diverse driving dataset for heterogeneous multitask learning, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2636–2645

2020
[12]

9523–9532

Y.Xu,X.Yang,L.Gong,H.-C.Lin,T.-Y.Wu,Y.Li,N.Vasconcelos, Explainable object-induced action decision for autonomous vehicles, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 9523–9532

2020
[13]

Y. Feng, W. Hua, Y. Sun, Nle-dm: Natural-language explanations for decision making of autonomous driving based on semantic scene understanding, IEEE Transactions on Intelligent Transportation Systems 24 (2023) 9780–9791

2023
[14]

Q. Wang, H. Hu, B. Yang, L. Song, C. Lv, Interpretable multi- task prediction neural network for autonomous vehicles, IEEE Transactions on Intelligent Transportation Systems (2025)

2025
[15]

S.Chowdhuri,T.Pankaj,K.Zipser, Multinet:Multi-modalmulti-task learning for autonomous driving, in: 2019 IEEE Winter Conference onApplicationsofComputerVision(WACV),IEEE,2019,pp.1496– 1504

2019
[16]

Bukowski, J

M. Bukowski, J. Kurek, I. Antoniuk, A. Jegorowa, Decision con- fidence assessment in multi-class classification, Sensors 21 (2021) 3834. MSH Azad et al.:Preprint submitted to ElsevierPage 11 of 12 Explainable Multi-Task Classification in Autonomous Vehicles

2021
[17]

J. Leo, J. Kalita, Incremental deep neural network learning using classification confidence thresholding, IEEE Transactions on Neural Networks and Learning Systems 33 (2021) 7706–7716

2021
[18]

A. A. Taha, L. Hennig, P. Knoth, Confidence estimation of classi- fication based on the distribution of the neural network output layer, arXiv preprint arXiv:2210.07745 (2022)

work page arXiv 2022
[19]

Z. Lv, W. Wang, K. Zhang, R. Tian, Y. Lv, M. Sun, Z. Xu, A high- confidenceinstanceboundaryregressionapproachanditsapplication in coal-gangue separation, Engineering Applications of Artificial Intelligence 132 (2024) 107894

2024
[20]

Thomas, P

J. Thomas, P. Mishra, D. M. Sharma, P. Krishnamurthy, Ltrc-iiith at ehrsql 2024: Enhancing reliability of text-to-sql systems through abstention and confidence thresholding, in: Proceedings of the 6th ClinicalNaturalLanguageProcessingWorkshop,2024,pp.697–702

2024
[21]

Z.Tang,K.V.Chuang,C.DeCarli,L.-W.Jin,L.Beckett,M.J.Keiser, B. N. Dugger, Interpretable classification of alzheimer’s disease pathologies with a convolutional neural network pipeline, Nature communications 10 (2019) 2173

2019
[22]

A. Wada, T. Akashi, G. Shih, A. Hagiwara, M. Nishizawa, Y.Hayakawa,J.Kikuta,K.Shimoji,K.Sano,K.Kamagata,etal.,Op- timizing gpt-4 turbo diagnostic accuracy in neuroradiology through promptengineeringandconfidencethresholds, Diagnostics14(2024) 1541

2024
[23]

Zhang, Y

J. Zhang, Y. Xie, G. Pang, Z. Liao, J. Verjans, W. Li, Z. Sun, J. He, Y.Li,C.Shen,etal., Viralpneumoniascreeningonchestx-raysusing confidence-aware anomaly detection, IEEE transactions on medical imaging 40 (2020) 879–890

2020
[24]

Thatikonda, M

M. Thatikonda, M. K. PK, F. Amsaad, A novel dynamic confidence threshold estimation ai algorithm for enhanced object detection, in: NAECON 2024-IEEE National Aerospace and Electronics Confer- ence, IEEE, 2024, pp. 359–363

2024
[25]

Cheng, Y

B. Cheng, Y. Wei, H. Shi, R. Feris, J. Xiong, T. Huang, Revisiting rcnn: On awakening the classification power of faster rcnn, in: ProceedingsoftheEuropeanconferenceoncomputervision(ECCV), 2018, pp. 453–468

2018
[26]

Van Ma, M

L. Van Ma, M. I. Hussain, J. Park, J. Kim, M. Jeon, Adaptive confi- dence threshold for bytetrack in multi-object tracking, in: 2023 12th International Conference on Control, Automation and Information Sciences (ICCAIS), IEEE, 2023, pp. 370–374

2023
[27]

Tambe, C

T. Tambe, C. Hooper, L. Pentecost, T. Jia, E.-Y. Yang, M. Donato, V. Sanh, P. Whatmough, A. M. Rush, D. Brooks, et al., Edgebert: Sentence-level energy optimizations for latency-aware multi-task nlp inference, in: MICRO-54: 54th Annual IEEE/ACM International Symposium on Microarchitecture, 2021, pp. 830–844

2021
[28]

Devlin, M.-W

J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies,volume1(longandshortpapers),2019,pp.4171–4186

2019
[29]

S. Qi, J. Li, Z. Sun, Adaptive confidence threshold algorithm for vehicle detection by employing temporal information, in: 2018 10th InternationalConferenceonIntelligentHuman-MachineSystemsand Cybernetics (IHMSC), volume 1, IEEE, 2018, pp. 348–352

2018
[30]

Gomez, G

A. Gomez, G. Diez, A. Salazar, A. Diaz, Animal identification in lowqualitycamera-trapimagesusingverydeepconvolutionalneural networks and confidence thresholds, in: International symposium on visual computing, Springer, 2016, pp. 747–756

2016
[31]

Bassani, M

D. Bassani, M. Reutlinger, H. Fischer, Leveraging machine learning predicted confidence for boosting assay submission and decision- making efficiencies, European Journal of Medicinal Chemistry 297 (2025) 117947

2025
[32]

J. Kim, A. Rohrbach, T. Darrell, J. Canny, Z. Akata, Textual expla- nations for self-driving vehicles, in: Proceedings of the European conference on computer vision (ECCV), 2018, pp. 563–578

2018
[33]

A2d2: Audi autonomous driving dataset.arXiv preprint arXiv:2004.06320, 2020

J. Geyer, Y. Kassahun, M. Mahmudi, X. Ricou, R. Durgesh, A. S. Chung,L.Hauswald,V.H.Pham,M.Mühlegg,S.Dorn,etal., A2d2: Audi autonomous driving dataset, arXiv preprint arXiv:2004.06320 (2020)

work page arXiv 2004
[34]

M. Gadd, D. De Martini, L. Marchegiani, P. Newman, L. Kunze, Sense–assess–explain(sax):Buildingtrustinautonomousvehiclesin challenging real-world driving scenarios, in: 2020 IEEE Intelligent Vehicles Symposium (IV), IEEE, 2020, pp. 150–155

2020
[35]

P.Sun,H.Kretzschmar,X.Dotiwalla,A.Chouard,V.Patnaik,P.Tsui, J. Guo, Y. Zhou, Y. Chai, B. Caine, et al., Scalability in perception forautonomousdriving:Waymoopendataset, in:Proceedingsofthe IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2446–2454

2020
[36]

Ramanishka, Y.-T

V. Ramanishka, Y.-T. Chen, T. Misu, K. Saenko, Toward driving sceneunderstanding:Adatasetforlearningdriverbehaviorandcausal reasoning, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 7699–7707

2018
[37]

PSI: A Benchmark for Human Interpretation and Response in Traffic Interactions

T.Chen,T.Jing,R.Tian,Y.Chen,J.Domeyer,H.Toyoda,R.Sherony, Z. Ding, Psi: A pedestrian behavior dataset for socially intelligent autonomous car, arXiv preprint arXiv:2112.02604 (2021)

work page internal anchor Pith review Pith/arXiv arXiv 2021
[38]

Cordts, M

M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be- nenson, U. Franke, S. Roth, B. Schiele, The cityscapes dataset for semantic urban scene understanding, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3213–3223

2016
[39]

A.Geiger,P.Lenz,C.Stiller,R.Urtasun, Visionmeetsrobotics:The kitti dataset, The international journal of robotics research 32 (2013) 1231–1237

2013
[40]

Caesar, V

H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A.Krishnan,Y.Pan,G.Baldan,O.Beijbom, nuscenes:Amultimodal dataset for autonomous driving, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11621–11631

2020
[41]

M. S. H. Azad, A. A. H. Imani, S. B. Shokouhi, Xai for transparent autonomous vehicles: A new approach to understanding decision- making in self-driving cars, in: 2024 14th International Conference onComputerandKnowledgeEngineering(ICCKE),IEEE,2024,pp. 194–199

2024
[42]

Q. Yuan, Y. Gao, J. Zhu, H. Xiong, Q. Xu, J. Wang, Summarizing vehicle driving decision-making methods on vulnerable road user collision avoidance, Digital Transportation and Safety 2 (2023) 23– 35

2023
[43]

Muslim, J

H. Muslim, J. Antona-Makoshi, A review of vehicle-to-vulnerable road user collisions on limited-access highways to support the devel- opment of automated vehicle safety assessments, Safety 8 (2022) 26

2022
[44]

Z. He, Y. Chen, B. King, L. Li, A configurable simulation frame- work for safety assessment of vulnerable road users, arXiv preprint arXiv:2510.19097 (2025). MSH Azad et al.:Preprint submitted to ElsevierPage 12 of 12

work page arXiv 2025

[1] [1]

J. Dong, S. Chen, S. Zong, T. Chen, S. Labi, Image transformer for explainable autonomous driving system, in: 2021 IEEE international intelligenttransportationsystemsconference(ITSC),IEEE,2021,pp. 2732–2737

2021

[2] [2]

H. Sun, M. Li, Z. Cui, Y. Huang, H. Chen, Semantic shapley-based counterfactual explanations for end-to-end autonomous driving, En- gineering Applications of Artificial Intelligence 159 (2025) 111638

2025

[3] [3]

J.Xie,Y.Zhang,Y.Qin,B.Wang,S.Dong,K.Li,Y.Xia, Ishuman- like decision making explainable? towards an explainable artificial intelligence for autonomous vehicles, Transportation Research In- terdisciplinary Perspectives 29 (2025) 101278

2025

[4] [4]

Albrecht, Ahuman-centricmethodforgeneratingcausalexplanations in natural language for autonomous vehicle motion planning, arXiv preprint arXiv:2206.08783 (2022)

B.Gyevnar,M.Tamborski,C.Wang,C.G.Lucas,S.B.Cohen,S.V. Albrecht, Ahuman-centricmethodforgeneratingcausalexplanations in natural language for autonomous vehicle motion planning, arXiv preprint arXiv:2206.08783 (2022)

work page arXiv 2022

[5] [5]

M. S. Hosseini Azad, S. Baradaran Shokouhi, A. A. Hamidi Imani, S. Atakishiyev, R. Goebel, An end-to-end decision-aware multi-scale attention-based model for explainable autonomous driving, 2026

2026

[6] [6]

C.Cao,X.Chen,J.Wang,Q.Song,R.Tan,Y.-H.Li,Sgdcl:Semantic- guided dynamic correlation learning for explainable autonomous driving, in: 33rd International Joint Conference on Artificial Intel- ligence (IJCAI 2024), International Joint Conferences on Artificial Intelligence, 2024, pp. 596–604

2024

[7] [7]

S. Meng, Y. Wang, Y. Cui, L.-P. Chau, Foundation model-assisted interpretable vehicle behavior decision making, Knowledge-Based Systems (2025) 113868

2025

[8] [8]

Z.Fu,K.Jiang,Y.Xu,Y.Wang,T.Wen,H.Gao,Z.Zhong,D.Yang, Top-down attention-based mechanisms for interpretable autonomous driving, IEEE Transactions on Intelligent Transportation Systems (2024)

2024

[9] [9]

D. Wang, C. Devin, Q.-Z. Cai, F. Yu, T. Darrell, Deep object-centric policies for autonomous driving, in: 2019 International Conference on Robotics and Automation (ICRA), IEEE, 2019, pp. 8853–8859

2019

[10] [10]

T. Jing, H. Xia, R. Tian, H. Ding, X. Luo, J. Domeyer, R. Sherony, Z. Ding, Inaction: Interpretable action decision making for au- tonomous driving, in: European Conference on Computer Vision, Springer, 2022, pp. 370–387

2022

[11] [11]

F. Yu, H. Chen, X. Wang, W. Xian, Y. Chen, F. Liu, V. Madhavan, T. Darrell, Bdd100k: A diverse driving dataset for heterogeneous multitask learning, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2636–2645

2020

[12] [12]

9523–9532

Y.Xu,X.Yang,L.Gong,H.-C.Lin,T.-Y.Wu,Y.Li,N.Vasconcelos, Explainable object-induced action decision for autonomous vehicles, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 9523–9532

2020

[13] [13]

Y. Feng, W. Hua, Y. Sun, Nle-dm: Natural-language explanations for decision making of autonomous driving based on semantic scene understanding, IEEE Transactions on Intelligent Transportation Systems 24 (2023) 9780–9791

2023

[14] [14]

Q. Wang, H. Hu, B. Yang, L. Song, C. Lv, Interpretable multi- task prediction neural network for autonomous vehicles, IEEE Transactions on Intelligent Transportation Systems (2025)

2025

[15] [15]

S.Chowdhuri,T.Pankaj,K.Zipser, Multinet:Multi-modalmulti-task learning for autonomous driving, in: 2019 IEEE Winter Conference onApplicationsofComputerVision(WACV),IEEE,2019,pp.1496– 1504

2019

[16] [16]

Bukowski, J

M. Bukowski, J. Kurek, I. Antoniuk, A. Jegorowa, Decision con- fidence assessment in multi-class classification, Sensors 21 (2021) 3834. MSH Azad et al.:Preprint submitted to ElsevierPage 11 of 12 Explainable Multi-Task Classification in Autonomous Vehicles

2021

[17] [17]

J. Leo, J. Kalita, Incremental deep neural network learning using classification confidence thresholding, IEEE Transactions on Neural Networks and Learning Systems 33 (2021) 7706–7716

2021

[18] [18]

A. A. Taha, L. Hennig, P. Knoth, Confidence estimation of classi- fication based on the distribution of the neural network output layer, arXiv preprint arXiv:2210.07745 (2022)

work page arXiv 2022

[19] [19]

Z. Lv, W. Wang, K. Zhang, R. Tian, Y. Lv, M. Sun, Z. Xu, A high- confidenceinstanceboundaryregressionapproachanditsapplication in coal-gangue separation, Engineering Applications of Artificial Intelligence 132 (2024) 107894

2024

[20] [20]

Thomas, P

J. Thomas, P. Mishra, D. M. Sharma, P. Krishnamurthy, Ltrc-iiith at ehrsql 2024: Enhancing reliability of text-to-sql systems through abstention and confidence thresholding, in: Proceedings of the 6th ClinicalNaturalLanguageProcessingWorkshop,2024,pp.697–702

2024

[21] [21]

Z.Tang,K.V.Chuang,C.DeCarli,L.-W.Jin,L.Beckett,M.J.Keiser, B. N. Dugger, Interpretable classification of alzheimer’s disease pathologies with a convolutional neural network pipeline, Nature communications 10 (2019) 2173

2019

[22] [22]

A. Wada, T. Akashi, G. Shih, A. Hagiwara, M. Nishizawa, Y.Hayakawa,J.Kikuta,K.Shimoji,K.Sano,K.Kamagata,etal.,Op- timizing gpt-4 turbo diagnostic accuracy in neuroradiology through promptengineeringandconfidencethresholds, Diagnostics14(2024) 1541

2024

[23] [23]

Zhang, Y

J. Zhang, Y. Xie, G. Pang, Z. Liao, J. Verjans, W. Li, Z. Sun, J. He, Y.Li,C.Shen,etal., Viralpneumoniascreeningonchestx-raysusing confidence-aware anomaly detection, IEEE transactions on medical imaging 40 (2020) 879–890

2020

[24] [24]

Thatikonda, M

M. Thatikonda, M. K. PK, F. Amsaad, A novel dynamic confidence threshold estimation ai algorithm for enhanced object detection, in: NAECON 2024-IEEE National Aerospace and Electronics Confer- ence, IEEE, 2024, pp. 359–363

2024

[25] [25]

Cheng, Y

B. Cheng, Y. Wei, H. Shi, R. Feris, J. Xiong, T. Huang, Revisiting rcnn: On awakening the classification power of faster rcnn, in: ProceedingsoftheEuropeanconferenceoncomputervision(ECCV), 2018, pp. 453–468

2018

[26] [26]

Van Ma, M

L. Van Ma, M. I. Hussain, J. Park, J. Kim, M. Jeon, Adaptive confi- dence threshold for bytetrack in multi-object tracking, in: 2023 12th International Conference on Control, Automation and Information Sciences (ICCAIS), IEEE, 2023, pp. 370–374

2023

[27] [27]

Tambe, C

T. Tambe, C. Hooper, L. Pentecost, T. Jia, E.-Y. Yang, M. Donato, V. Sanh, P. Whatmough, A. M. Rush, D. Brooks, et al., Edgebert: Sentence-level energy optimizations for latency-aware multi-task nlp inference, in: MICRO-54: 54th Annual IEEE/ACM International Symposium on Microarchitecture, 2021, pp. 830–844

2021

[28] [28]

Devlin, M.-W

J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies,volume1(longandshortpapers),2019,pp.4171–4186

2019

[29] [29]

S. Qi, J. Li, Z. Sun, Adaptive confidence threshold algorithm for vehicle detection by employing temporal information, in: 2018 10th InternationalConferenceonIntelligentHuman-MachineSystemsand Cybernetics (IHMSC), volume 1, IEEE, 2018, pp. 348–352

2018

[30] [30]

Gomez, G

A. Gomez, G. Diez, A. Salazar, A. Diaz, Animal identification in lowqualitycamera-trapimagesusingverydeepconvolutionalneural networks and confidence thresholds, in: International symposium on visual computing, Springer, 2016, pp. 747–756

2016

[31] [31]

Bassani, M

D. Bassani, M. Reutlinger, H. Fischer, Leveraging machine learning predicted confidence for boosting assay submission and decision- making efficiencies, European Journal of Medicinal Chemistry 297 (2025) 117947

2025

[32] [32]

J. Kim, A. Rohrbach, T. Darrell, J. Canny, Z. Akata, Textual expla- nations for self-driving vehicles, in: Proceedings of the European conference on computer vision (ECCV), 2018, pp. 563–578

2018

[33] [33]

A2d2: Audi autonomous driving dataset.arXiv preprint arXiv:2004.06320, 2020

J. Geyer, Y. Kassahun, M. Mahmudi, X. Ricou, R. Durgesh, A. S. Chung,L.Hauswald,V.H.Pham,M.Mühlegg,S.Dorn,etal., A2d2: Audi autonomous driving dataset, arXiv preprint arXiv:2004.06320 (2020)

work page arXiv 2004

[34] [34]

M. Gadd, D. De Martini, L. Marchegiani, P. Newman, L. Kunze, Sense–assess–explain(sax):Buildingtrustinautonomousvehiclesin challenging real-world driving scenarios, in: 2020 IEEE Intelligent Vehicles Symposium (IV), IEEE, 2020, pp. 150–155

2020

[35] [35]

P.Sun,H.Kretzschmar,X.Dotiwalla,A.Chouard,V.Patnaik,P.Tsui, J. Guo, Y. Zhou, Y. Chai, B. Caine, et al., Scalability in perception forautonomousdriving:Waymoopendataset, in:Proceedingsofthe IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2446–2454

2020

[36] [36]

Ramanishka, Y.-T

V. Ramanishka, Y.-T. Chen, T. Misu, K. Saenko, Toward driving sceneunderstanding:Adatasetforlearningdriverbehaviorandcausal reasoning, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 7699–7707

2018

[37] [37]

PSI: A Benchmark for Human Interpretation and Response in Traffic Interactions

T.Chen,T.Jing,R.Tian,Y.Chen,J.Domeyer,H.Toyoda,R.Sherony, Z. Ding, Psi: A pedestrian behavior dataset for socially intelligent autonomous car, arXiv preprint arXiv:2112.02604 (2021)

work page internal anchor Pith review Pith/arXiv arXiv 2021

[38] [38]

Cordts, M

M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be- nenson, U. Franke, S. Roth, B. Schiele, The cityscapes dataset for semantic urban scene understanding, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3213–3223

2016

[39] [39]

A.Geiger,P.Lenz,C.Stiller,R.Urtasun, Visionmeetsrobotics:The kitti dataset, The international journal of robotics research 32 (2013) 1231–1237

2013

[40] [40]

Caesar, V

H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, A.Krishnan,Y.Pan,G.Baldan,O.Beijbom, nuscenes:Amultimodal dataset for autonomous driving, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11621–11631

2020

[41] [41]

M. S. H. Azad, A. A. H. Imani, S. B. Shokouhi, Xai for transparent autonomous vehicles: A new approach to understanding decision- making in self-driving cars, in: 2024 14th International Conference onComputerandKnowledgeEngineering(ICCKE),IEEE,2024,pp. 194–199

2024

[42] [42]

Q. Yuan, Y. Gao, J. Zhu, H. Xiong, Q. Xu, J. Wang, Summarizing vehicle driving decision-making methods on vulnerable road user collision avoidance, Digital Transportation and Safety 2 (2023) 23– 35

2023

[43] [43]

Muslim, J

H. Muslim, J. Antona-Makoshi, A review of vehicle-to-vulnerable road user collisions on limited-access highways to support the devel- opment of automated vehicle safety assessments, Safety 8 (2022) 26

2022

[44] [44]

Z. He, Y. Chen, B. King, L. Li, A configurable simulation frame- work for safety assessment of vulnerable road users, arXiv preprint arXiv:2510.19097 (2025). MSH Azad et al.:Preprint submitted to ElsevierPage 12 of 12

work page arXiv 2025