Pith. sign in

REVIEW 5 major objections 6 minor 55 references

The paper organizes explainable-AI methods for object detection into four mechanism-based families and benchmarks them, finding FSOD and BSED lead on COCO and VOC.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A survey organizing explainable AI methods for object detection into perturbation, gradient, backpropagation, and graph based families, with an overview of datasets, metrics, and publication trends.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A readable, accurate entry-level survey of XAI for object detection, but the 'comprehensive' framing overreaches and the selection protocol is absent; conditional acceptance at best. the 5 major comments →

arxiv 2509.01991 v1 pith:U3UNCZLW submitted 2025-09-02 cs.CV

Explaining What Machines See: XAI Strategies in Deep Object Detection Models

FatemehSadat Seyedmomeni , Mohammad Ali Keyvanrad This is my paper
classification cs.CV
keywords Explainable AIObject detectionSaliency mapsPerturbation-based methodsGradient-based methodsBackpropagation-based attributionGraph-based explainabilityFidelity metrics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This review argues that explainability methods for object detection can be organized into four mechanism-based families: perturbation/occlusion, gradient-based, backpropagation-based, and graph-based approaches. It catalogs more than twenty named methods, maps them onto detector architectures such as YOLO and Faster R-CNN, and compiles the datasets and evaluation metrics used to judge them. Using published comparisons on COCO and PASCAL VOC, it highlights FSOD and BSED as particularly strong on insertion, deletion, and energy-based pointing game metrics. The paper intends to give practitioners a structured basis for choosing an explanation method, filling a perceived gap left by classification-focused XAI reviews.

Core claim

The paper's central claim is that the explainable-object-detection literature, though fragmented, can be systematically understood through a four-category taxonomy based on underlying mechanism, and that within this taxonomy a handful of recent methods—notably FSOD and BSED—set the current standard for fidelity. It also establishes that research activity in this niche has sharply accelerated from 2022 to mid-2025, and that evaluation practice is converging on a small set of fidelity metrics (Insertion, Deletion, EPG) applied on COCO and PASCAL VOC. The authors treat these regularities as evidence that explanation for detection is becoming a mature subfield with identifiable best practices.

What carries the argument

The organizing machinery is the mechanism-based taxonomy: perturbation/occlusion methods (random or hierarchical masking that measures output change), gradient-based methods (derivatives of detection output with respect to features or input), backpropagation-based attribution (relevance propagation backward through the network), and graph-based methods (structured relationships among regions or features). Within this taxonomy, the comparative evaluation relies on three fidelity metrics—Insertion, Deletion, and Explanation-based Pointing Game (EPG)—which the review uses to benchmark methods on the COCO and VOC datasets.

Load-bearing premise

The review's claim of comprehensive coverage rests on an unstated assumption that the subset of methods and papers it examines in detail is representative of the full body of XAI-for-object-detection research; the paper says only 'a subset has been thoroughly examined' but never specifies how that subset was selected.

What would settle it

A systematic literature search that identifies a peer-reviewed explainability method for object detection whose mechanism does not fit any of the four categories (perturbation, gradient, backpropagation, graph) would refute the taxonomy's completeness; alternatively, re-running the reported COCO/VOC comparisons with a single unified evaluation protocol and showing that FSOD/BSED no longer lead would refute the review's comparative conclusions.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners can use the taxonomy to pre-filter explanation techniques by whether their detector is single-stage (YOLO/SSD) or two-stage (Faster R-CNN) and by their need for model-agnostic versus internal access.
  • FSOD and BSED, by virtue of their leading scores, become the default first-choice baselines for new explainable-detection work targeting fidelity on COCO/VOC.
  • The convergence on Insertion/Deletion/EPG metrics enables cross-paper comparison, though the review notes differences in experimental setups currently prevent definitive rankings.
  • The accelerating publication trend (roughly a four-fold jump in monthly output from 2022 to 2025) suggests that framework gaps—evaluation standardization, real-time explainers, human alignment—are active research fronts.
  • The D-Deletion and D-Minimum Subset metrics extend fidelity evaluation to the multi-instance, localization-aware setting of detection.

Where Pith is reading between the lines

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

  • The four-way taxonomy may underplay hybrid methods: several surveyed techniques (e.g., WS-Grad, Concat-Grad) combine gradient-based components, suggesting that a composite or dimension-based taxonomy (model access, computational budget, output granularity) might predict method choice better than mechanism alone.
  • If the reported FSOD/BSED advantages replicate under controlled, matched training conditions across detectors, the field could shift toward learned explainers that amortize Shapley-value computation, making real-time explanation practical.
  • The review's selection of datasets (COCO, VOC, CrowdHuman, BDD-100K, KITTI, GRAZPEDWRI-DX) implies a claim that explanation quality must be validated across scene complexity, density, and domain shift; a testable corollary is that methods ranking high on COCO may not retain their lead on dense or medical datasets like CrowdHuman or GRAZPEDWRI-DX.
  • Standardized benchmark protocols for XAI in detection (fixed masks, fixed baselines) would likely shrink reported performance gaps between methods—the current gaps may partly reflect evaluation variability across papers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper is a review of explainable AI (XAI) methods for object detection. It proposes a four-category taxonomy (perturbation-, gradient-, backpropagation-, and graph-based), describes representative methods (D-RISE, BODEM, D-CLOSE, FSOD, G-CAME, SS-GradCAM, L-CRP, AOG, SRR, etc.), reviews datasets and evaluation metrics, reports publication-trend statistics from Google Scholar (2022–mid-2025), and presents comparative tables on MS-COCO and PASCAL VOC. The abstract claims a comprehensive analysis; Section 1 acknowledges that only a subset of the literature was examined.

Significance. If taken as a curated introduction rather than a systematic review, the paper has pedagogical value: the method descriptions are mostly accurate, the summary tables condense key ideas, and the comparative tables are explicitly attributed to the original papers [9], [35]. The taxonomy and the trend data are the main claimed contributions. However, the 'comprehensive' claim and the comparative conclusions go beyond what the evidence supports, so the paper's significance depends on substantial revision of scope and claims.

major comments (5)
  1. [Abstract & Section 1 (last two paragraphs)] The central claim of a 'comprehensive analysis' is not supported because no review methodology is provided. The paper states that 'only a subset has been thoroughly examined' but does not specify how that subset was selected, which databases were queried beyond an unspecified Google Scholar search, which inclusion/exclusion criteria were applied, or how screening was performed. The resulting taxonomy (Figure 20) contains only two graph-based methods, one backpropagation-based method, and one 'other' method; this may be an artifact of the selection rather than a faithful map of the field. The claim should either be supported by a systematic protocol or replaced with an explicitly non-exhaustive scope.
  2. [Section 5, Tables 8 and 9, and text after 'Based on the results gathered'] The relative-performance conclusions (e.g., 'FSOD demonstrates superior performance', 'BSED ... optimal performance in the Deletion metric') are taken directly from [9] and [35], which propose FSOD and BSED, using a random 10% subset of the validation set with no confidence intervals or significance tests. The paper itself acknowledges that 'differences in experimental design prevent providing a definitive conclusion', yet the following sentences present the numbers as though they establish superiority. These should be reported as results from the cited studies, attributed accordingly, and the over-strong language should be removed.
  3. [Section 1, 'To investigate the trends ...' through Figure 2] The publication-trend statistics are not reproducible. The exact Google Scholar query string, the search date, deduplication settings, language/venue filters, and the procedure for assigning articles to years are not stated. Since 'statistical analysis of publication trends' is one of the paper's stated contributions, the numbers should be backed by a reproducible search (e.g., exact query and access date, preferably with a second database such as Scopus or OpenAlex) or removed/clearly labeled as rough, non-reproducible estimates.
  4. [Sections 1.2 and 2 (overall scope)] The review claims to cover 'state-of-the-art explainability methods specifically applied to object detection models', yet it restricts architectural coverage to YOLO, SSD, Faster R-CNN, and EfficientDet, with no discussion of transformer-based detectors (e.g., DETR, Deformable DETR) or explainability methods developed for them. Likewise, causal XAI and attention-based explanation techniques are absent. If the intended scope is CNN-based one-/two-stage detectors, that limitation should be stated explicitly, and the 'comprehensive' claim should be revised accordingly. Otherwise, these families need to be represented in the taxonomy.
  5. [Sections 1.3.3 and 2.3, Table 3] There is an internal inconsistency in the backpropagation-based category. Section 1.3.3 discusses LRP and CRP as prominent examples, and Table 8 evaluates CRP, but Section 2.3 and Table 3 cover only L-CRP. No object-detection-specific treatment of LRP or CRP is provided. This gap should be filled or explicitly noted as outside the selected subset. Without this, the taxonomy's backpropagation branch is not comparable to the other branches in terms of coverage.
minor comments (6)
  1. [Figures] Figure numbering is inconsistent: there are duplicate numbers (Figure 28 appears twice, Figure 30 appears twice) and Figure 29 appears after Figure 28. Please renumber all figures.
  2. [References] Several references are duplicated: [18] and [32] are the same Petsiuk et al. paper; [19] and [29] are the same Singh et al. paper. Consolidate duplicates.
  3. [Section 5, Tables 8-9] The metric is introduced as EBPG (Energy-Based Pointing Game) in Section 4.1.8 but appears as EPG in Section 5 and the tables. Use one consistent abbreviation and define it where first used.
  4. [Section 1.2] Subsection numbering is confused: 'Two-Stage Models' appears as 1.1.1 after Section 1.2, but it should be a subsection of 1.2. Renumber 1.1.1/1.1.2/1.2.1 etc. accordingly.
  5. [Figure 10 caption] Typo: 'Framwork' should be 'Framework'.
  6. [Section 3] The dataset descriptions are informative, but there is no table mapping which XAI methods were evaluated on which datasets. Such a mapping would substantially increase the practical value of the review.

Circularity Check

0 steps flagged

No significant circularity: this is a descriptive review with no derived predictions; comparative numbers are attributed to external sources.

full rationale

The paper is a narrative review and taxonomy of explainability methods for object detection. It makes no first-principles derivation, fits no parameters, and reports no new empirical predictions. The publication-trend statistics are direct Google Scholar keyword counts, not derived from the methods under review. The comparative tables in Section 5 are explicitly sourced to the external FSOD and BSED papers ([9], [35]), and the review itself notes that differences in experimental design prevent a definitive conclusion; it does not present these as its own fitted or derived results. The taxonomy is a categorization scheme, and while the representativeness of the reviewed subset is a legitimate coverage limitation (Section 1: 'only a subset has been thoroughly examined'), that is a selection-bias concern, not circularity. No equation is defined in terms of the result it is used to support, and no load-bearing argument reduces to a self-citation by the present authors. Accordingly, no circular step is identified.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The review introduces no fitted parameters and no invented entities. Its substantive assumptions are editorial: that the four-way taxonomy is the right organizing scheme, that Google Scholar keyword counts measure research interest, that the evaluation numbers reproduced from [9] and [35] are valid and comparable, and that the subset of papers given detailed treatment is representative of the field. The last assumption is load-bearing for the 'comprehensive' claim and is not supported by a stated selection methodology.

axioms (4)
  • domain assumption The four-way taxonomy (perturbation, gradient, backpropagation, graph) is the natural organizing scheme for XAI in object detection.
    Section 2 and Figure 20; the paper asserts this categorization rather than deriving it.
  • domain assumption Google Scholar keyword counts with the two specified query pairs measure research interest in explainable object detection.
    Section 1; counts are approximate, and search date and deduplication method are not given.
  • domain assumption The evaluation numbers reproduced from [9] and [35] are valid and comparable across methods.
    Tables 8 and 9; the paper relies on two cited papers (same author group) for all comparative numbers.
  • domain assumption The reviewed subset of methods is representative of the field.
    Section 1 states only a subset was thoroughly examined; selection criteria are not specified.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Explaining What Machines See: XAI Strategies in Deep Object Detection Models." pith.science (2026). https://pith.science/paper/U3UNCZLW

@misc{pith2026250901991,
  author       = {Pith},
  title        = {Pith review of: Explaining What Machines See: XAI Strategies in Deep Object Detection Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U3UNCZLW}},
  note         = {Machine review of arXiv:2509.01991}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In recent years, deep learning has achieved unprecedented success in various computer vision tasks, particularly in object detection. However, the black-box nature and high complexity of deep neural networks pose significant challenges for interpretability, especially in critical domains such as autonomous driving, medical imaging, and security systems. Explainable Artificial Intelligence (XAI) aims to address this challenge by providing tools and methods to make model decisions more transparent, interpretable, and trust-worthy for humans. This review provides a comprehensive analysis of state-of-the-art explain-ability methods specifically applied to object detection models. The paper be-gins by categorizing existing XAI techniques based on their underlying mechanisms-perturbation-based, gradient-based, backpropagation-based, and graph-based methods. Notable methods such as D-RISE, BODEM, D-CLOSE, and FSOD are discussed in detail. Furthermore, the paper investigates their applicability to various object detection architectures, including YOLO, SSD, Faster R-CNN, and EfficientDet. Statistical analysis of publication trends from 2022 to mid-2025 shows an accelerating interest in explainable object detection, indicating its increasing importance. The study also explores common datasets and evaluation metrics, and highlights the major challenges associated with model interpretability. By providing a structured taxonomy and a critical assessment of existing methods, this review aims to guide researchers and practitioners in selecting suitable explainability techniques for object detection applications and to foster the development of more interpretable AI systems.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    Deep learning,

    Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, May 2015, doi: 10.1038/nature14539

    doi:10.1038/nature14539 2015

  2. [2]

    ‘Why Should I Trust You?’ | Proceedings of the 22nd ACM SIGKDD Interna- tional Conference on Knowledge Discovery and Data Mining

    “‘Why Should I Trust You?’ | Proceedings of the 22nd ACM SIGKDD Interna- tional Conference on Knowledge Discovery and Data Mining.” Accessed: Jun. 05, 2025. [Online]. Available: https://dl.acm.org/doi/abs/10.1145/2939672.2939778

    arXiv 2025

  3. [3]

    (PDF) Causability and explainabilty of artificial intelligence in medicine,

    “(PDF) Causability and explainabilty of artificial intelligence in medicine,” Re- searchGate, Dec. 2024, doi: 10.1002/widm.1312

    doi:10.1002/widm.1312 2024

  4. [4]

    DARPA’s explainable artificial intelligence program,

    G. D. A. DW, “DARPA’s explainable artificial intelligence program,” AI Mag, vol. 40, no. 2, p. 44, 2019

    work page 2019

  5. [5]

    Opportunities and Challenges in Explainable Artificial Intelligence (XAI): A Survey | Request PDF

    “Opportunities and Challenges in Explainable Artificial Intelligence (XAI): A Survey | Request PDF.” Accessed: Jun. 05, 2025. [Online]. Available: https://www.researchgate.net/publication/342377682_Opportunities_and_Chal- lenges_in_Explainable_Artificial_Intelligence_XAI_A_Survey

    work page arXiv 2025

  6. [6]

    Explainable AI: A comprehensive review of the main methods,

    F. Dallanoce, “Explainable AI: A comprehensive review of the main methods,” MLearning Ai, 2022. 68 F.SeyedMomeni et al

    work page 2022

  7. [7]

    What is an attention mechanism? | IBM

    “What is an attention mechanism? | IBM.” Accessed: May 31, 2025. [Online]. Available: https://www.ibm.com/think/topics/attention-mechanism

    work page 2025

  8. [8]

    Explainable AI: The Path To Human -Friendly Artificial Intelli- gence,

    CodeTrade, “Explainable AI: The Path To Human -Friendly Artificial Intelli- gence,” Custom Software Development Company. Accessed: Jun. 02, 2025. [Online]. Available: https://www.codetrade.io/blog/explainable -ai-the-path-to- human-friendly-artificial-intelligence/

    work page 2025

  9. [9]

    Fast Explanation Using Shapley Value for Object Detection,

    M. Kuroki and T. Yamasaki, “Fast Explanation Using Shapley Value for Object Detection,” IEEE Access , vol. 12, pp. 31047 –31054, 2024, doi: 10.1109/ACCESS.2024.3369890

    work page arXiv 2024

  10. [10]

    Object detection with deep learning: A review,

    Z.-Q. Zhao, P. Zheng, S. Xu, and X. Wu, “Object detection with deep learning: A review,” IEEE Trans. Neural Netw. Learn. Syst. , vol. 30, no. 11, pp. 3212 – 3232, 2019

    work page 2019

  11. [11]

    (PDF) Object Detection Performance: A Comparative Study

    “(PDF) Object Detection Performance: A Comparative Study.” Accessed: May 25, 2025. [Online]. Available: https://www.researchgate.net/publica- tion/372570082_Object_Detection_Performance_A_Comparative_Study

    work page arXiv 2025

  12. [12]

    Rich feature hierarchies for accurate object detection and semantic segmentation,

    R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hierarchies for accurate object detection and semantic segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2014, pp. 580 – 587

    work page 2014

  13. [13]

    Fast R -CNN,

    R. Girshick, “Fast R -CNN,” presented at the Proceedings of the IEEE Interna- tional Conference on Computer Vision, 2015, pp. 1440–1448. Accessed: Jun. 03,

    work page 2015

  14. [14]

    Faster R -CNN: Towards Real -Time Object Detection with Region Proposal Networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster R -CNN: Towards Real -Time Object Detection with Region Proposal Networks,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 39, no. 6, pp. 1137 –1149, Jun. 2017, doi: 10.1109/TPAMI.2016.2577031

    arXiv 2017

  15. [15]

    You only look once: Uni- fied, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Uni- fied, real-time object detection,” in Proceedings of the IEEE conference on com- puter vision and pattern recognition, 2016, pp. 779–788

    work page 2016

  16. [16]

    SSD: Single Shot MultiBox Detector,

    W. Liu et al., “SSD: Single Shot MultiBox Detector,” vol. 9905, 2016, pp. 21 –

    work page 2016

  17. [17]

    Efficientdet: Scalable and efficient object de- tection,

    M. Tan, R. Pang, and Q. V. Le, “Efficientdet: Scalable and efficient object de- tection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10781–10790

    work page 2020

  18. [20]

    GSM -HM: Generation of Saliency Maps for Black-Box Object Detection Model Based on Hierarchical Masking,

    Y. Yan, X. Li, Y. Zhan, L. Sun, and J. Zhu, “GSM -HM: Generation of Saliency Maps for Black-Box Object Detection Model Based on Hierarchical Masking,” IEEE Access , vol. 10, pp. 98268 –98277, 2022, doi: 10.1109/ACCESS.2022.3206379. 69

    work page arXiv 2022

  19. [21]

    A unified approach to interpreting model predic- tions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predic- tions,” Adv. Neural Inf. Process. Syst., vol. 30, 2017

    work page 2017

  20. [22]

    (PDF) RISE: Randomized Input Sampling for Explanation of Black -box Mod- els

    “(PDF) RISE: Randomized Input Sampling for Explanation of Black -box Mod- els.” Accessed: Jun. 05, 2025. [Online]. Available: https://www.re- searchgate.net/publica- tion/325893765_RISE_Randomized_Input_Sampling_for_Explanation_of_Bla ck-box_Models

    work page arXiv 2025

  21. [23]

    MFPP: Morphological Fragmental Perturbation Pyramid for Black -Box Model Explanations,

    Q. Yang, X. Zhu, J.-K. Fwu, Y. Ye, G. You, and Y. Zhu, “MFPP: Morphological Fragmental Perturbation Pyramid for Black -Box Model Explanations,” in 2020 25th International Conference on Pattern Recognition (ICPR) , Jan. 2021, pp. 1376–1383. doi: 10.1109/ICPR48806.2021.9413046

    arXiv 2020

  22. [24]

    Deep inside convolutional net- works: Visualising image classification models and saliency maps,

    K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside convolutional net- works: Visualising image classification models and saliency maps,” ArXiv Prepr. ArXiv13126034, 2013

    work page 2013

  23. [25]

    Learning deep features for discriminative localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2921–2929

    work page 2016

  24. [26]

    Grad-CAM: Visual Explanations from Deep Networks via Gradient -based Lo- calization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual Explanations from Deep Networks via Gradient -based Lo- calization,” Int. J. Comput. Vis. , vol. 128, no. 2, pp. 336 –359, Feb. 2020, doi: 10.1007/s11263-019-01228-7

    doi:10.1007/s11263-019-01228-7 2020

  25. [27]

    GradCAM-architecture.png (PNG Image, 1600 × 764 pixels) — Scaled (95%)

    “GradCAM-architecture.png (PNG Image, 1600 × 764 pixels) — Scaled (95%).” Accessed: May 12, 2025. [Online]. Available: https://learnopencv.com/wp -con- tent/uploads/2023/12/GradCAM-architecture.png

    work page 2025

  26. [28]

    Grad - CAM++: Generalized Gradient -Based Visual Explanations for Deep Convolu- tional Networks,

    A. Chattopadhay, A. Sarkar, P. Howlader, and V. N. Balasubramanian, “Grad - CAM++: Generalized Gradient -Based Visual Explanations for Deep Convolu- tional Networks,” in 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Mar. 2018, pp. 839–847. doi: 10.1109/WACV.2018.00097

    arXiv 2018

  27. [29]

    Explainable deep learning models in medical image analysis,

    A. Singh, S. Sengupta, and V. Lakshminarayanan, “Explainable deep learning models in medical image analysis,” J. Imaging, vol. 6, no. 6, p. 52, 2020

    work page 2020

  28. [30]

    Heatmap -based Expla- nation of YOLOv5 Object Detection with Layer -wise Relevance Propagation,

    A. Karasmanoglou, M. Antonakakis, and M. Zervakis, “Heatmap -based Expla- nation of YOLOv5 Object Detection with Layer -wise Relevance Propagation,” in 2022 IEEE International Conference on Imaging Systems and Techniques (IST), Jun. 2022, pp. 1–6. doi: 10.1109/IST55454.2022.9827744

    work page arXiv 2022

  29. [31]

    Contrastive Relevance Propagation for Interpreting Predictions by a Single -Shot Object De- tector,

    H. Tsunakawa, Y. Kameya, H. Lee, Y. Shinya, and N. Mitsumoto, “Contrastive Relevance Propagation for Interpreting Predictions by a Single -Shot Object De- tector,” in 2019 International Joint Conference on Neural Networks (IJCNN) , Jul. 2019, pp. 1–9. doi: 10.1109/IJCNN.2019.8851770

    work page arXiv 2019

  30. [32]

    Black-box explanation of object detectors via saliency maps,

    V. Petsiuk et al., “Black-box explanation of object detectors via saliency maps,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 11443–11452

    work page 2021

  31. [33]

    Surrogate Object Detection Ex- plainer (SODEx) with YOLOv4 and LIME,

    J. H. Sejr, P. Schneider -Kamp, and N. Ayoub, “Surrogate Object Detection Ex- plainer (SODEx) with YOLOv4 and LIME,” Mach. Learn. Knowl. Extr., vol. 3, no. 3, Art. no. 3, Sep. 2021, doi: 10.3390/make3030033. 70 F.SeyedMomeni et al

    work page doi:10.3390/make3030033 2021

  32. [34]

    Model -agnostic explainable artificial intelligence for object detection in image data,

    M. Moradi, K. Yan, D. Colwell, M. Samwald, and R. Asgari, “Model -agnostic explainable artificial intelligence for object detection in image data,” Eng. Appl. Artif. Intell. , vol. 137, p. 109183, Nov. 2024, doi: 10.1016/j.en- gappai.2024.109183

    arXiv 2024

  33. [35]

    BSED: Baseline Shapley -Based Explainable De- tector,

    M. Kuroki and T. Yamasaki, “BSED: Baseline Shapley -Based Explainable De- tector,” IEEE Access , vol. 12, pp. 57959 –57973, 2024, doi: 10.1109/ACCESS.2024.3391424

    work page arXiv 2024

  34. [36]

    Towards Better Explanations for Object Detection,

    V. B. Truong, T. T. H. Nguyen, V. T. K. Nguyen, Q. K. Nguyen, and Q. H. Cao, “Towards Better Explanations for Object Detection,” in Proceedings of the 15th Asian Conference on Machine Learning, PMLR, Feb. 2024, pp. 1385–1400. Ac- cessed: Jun. 05, 2025. [Online]. Available: https://proceed- ings.mlr.press/v222/truong24a.html

    work page 2024

  35. [37]

    doi: 10.1007/978-3-319-46448-0_2

    doi:10.1007/978-3-319-46448-0_2

  36. [38]

    Ex- plain to Fix: A Framework to Interpret and Correct DNN Object Detector Pre- dictions,

    D. Gudovskiy, A. Hodgkinson, T. Yamaguchi, Y. Ishii, and S. Tsukizawa, “Ex- plain to Fix: A Framework to Interpret and Correct DNN Object Detector Pre- dictions,” Nov. 19, 2018, arXiv: arXiv:1811.08011. doi: 10.48550/arXiv.1811.08011

    Pith/arXiv arXiv doi:10.48550/arxiv.1811.08011 2018

  37. [39]

    On the black -box ex- plainability of object detection models for safe and trustworthy industrial appli- cations,

    A. Andres, A. Martinez -Seras, I. Laña, and J. Del Ser, “On the black -box ex- plainability of object detection models for safe and trustworthy industrial appli- cations,” Results Eng. , vol. 24, p. 103498, Dec. 2024, doi: 10.1016/j.rineng.2024.103498

    work page arXiv 2024

  38. [40]

    G-CAME: Gaussian-Class Activation Mapping Explainer for Object Detectors

    Q. K. Nguyen, T. T. H. Nguyen, V. T. K. Nguyen, V. B. Truong, and Q. H. Cao, “G-CAME: Gaussian -Class Activation Mapping Explainer for Object Detec- tors,” Jun. 06, 2023, arXiv: arXiv:2306.03400. doi: 10.48550/arXiv.2306.03400

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2306.03400 2023

  39. [41]

    Human attention guided explain- able artificial intelligence for computer vision models,

    G. Liu, J. Zhang, A. B. Chan, and J. H. Hsiao, “Human attention guided explain- able artificial intelligence for computer vision models,” Neural Netw., vol. 177, p. 106392, 2024

    work page 2024

  40. [42]

    Spatial Sensitive Grad -CAM++: Improved Visual Explanation for Object Detectors via Weighted Combination of Gradient Map

    T. Yamauchi, “Spatial Sensitive Grad -CAM++: Improved Visual Explanation for Object Detectors via Weighted Combination of Gradient Map”

    work page

  41. [43]

    Spatial sensitive grad-cam: Visual explanations for object detection by incorporating spatial sensitivity,

    T. Yamauchi and M. Ishikawa, “Spatial sensitive grad-cam: Visual explanations for object detection by incorporating spatial sensitivity,” in 2022 IEEE Interna- tional Conference on Image Processing (ICIP), IEEE, 2022, pp. 256–260

    work page 2022

  42. [44]

    Optimisation of Deep Learning Small-Object Detectors with Novel Explainable Verification

    “Optimisation of Deep Learning Small-Object Detectors with Novel Explainable Verification.” Accessed: May 12, 2025. [Online]. Available: https://www.mdpi.com/1424-8220/22/15/5596

    work page 2025

  43. [45]

    Gradient-Based Instance-Specific Visual Explanations for Object Specification and Object Discrimination,

    C. Zhao, J. H. Hsiao, and A. B. Chan, “Gradient-Based Instance-Specific Visual Explanations for Object Specification and Object Discrimination,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 46, no. 9, pp. 5967 –5985, Sep. 2024, doi: 10.1109/TPAMI.2024.3380604

    work page arXiv 2024

  44. [46]

    Towards Interpretable Object Detection by Unfolding La- tent Structures,

    T. Wu and X. Song, “Towards Interpretable Object Detection by Unfolding La- tent Structures,” in 2019 IEEE/CVF International Conference on Computer 71 Vision (ICCV) , Seoul, Korea (South): IEEE, Oct. 2019, pp. 6032 –6042. doi: 10.1109/ICCV.2019.00613

    work page arXiv 2019

  45. [47]

    Revealing hidden context bias in segmentation and object detection through concept -spe- cific explanations,

    M. Dreyer, R. Achtibat, T. Wiegand, W. Samek, and S. Lapuschkin, “Revealing hidden context bias in segmentation and object detection through concept -spe- cific explanations,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3829–3839

    work page 2023

  46. [48]

    Intrinsic Explainability for End -to-End Object Detection,

    L. Fernandes, J. N. D. Fernandes, M. Calado, J. R. Pinto, R. Cerqueira, and J. S. Cardoso, “Intrinsic Explainability for End -to-End Object Detection,” IEEE Ac- cess, vol. 12, pp. 2623–2634, 2024, doi: 10.1109/ACCESS.2023.3347038

    arXiv 2024

  47. [49]

    Towards Human -Like Interpretable Object Detection Via Spatial Relation Encoding,

    J. U. Kim, S. Park, and Y. M. Ro, “Towards Human -Like Interpretable Object Detection Via Spatial Relation Encoding,” in 2020 IEEE International Confer- ence on Image Processing (ICIP) , Oct. 2020, pp. 3284 –3288. doi: 10.1109/ICIP40778.2020.9190724

    arXiv 2020

  48. [50]

    The Pascal Visual Object Classes Challenge: A Retrospective | International Journal of Computer Vision

    “The Pascal Visual Object Classes Challenge: A Retrospective | International Journal of Computer Vision.” Accessed: May 24, 2025. [Online]. Available: https://link.springer.com/article/10.1007/s11263-014-0733-5

    doi:10.1007/s11263-014-0733-5 2025

  49. [51]

    Microsoft coco: Common objects in context,

    T.-Y. Lin et al., “Microsoft coco: Common objects in context,” in Computer vi- sion–ECCV 2014: 13th European conference, zurich, Switzerland, September 6- 12, 2014, proceedings, part v 13, Springer, 2014, pp. 740–755

    work page 2014

  50. [52]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning,

    F. Yu et al., “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

    work page 2020

  51. [53]

    CrowdHuman: A Benchmark for Detecting Human in a Crowd,

    S. Shao et al., “CrowdHuman: A Benchmark for Detecting Human in a Crowd,” Apr. 30, 2018, arXiv: arXiv:1805.00123. doi: 10.48550/arXiv.1805.00123

    Pith/arXiv arXiv doi:10.48550/arxiv.1805.00123 2018

  52. [54]

    A pediatric wrist trauma X -ray dataset (GRAZPEDWRI -DX) for machine learning,

    E. Nagy, M. Janisch, F. Hržić, E. Sorantin, and S. Tschauner, “A pediatric wrist trauma X -ray dataset (GRAZPEDWRI -DX) for machine learning,” Sci. Data, vol. 9, no. 1, p. 222, May 2022, doi: 10.1038/s41597-022-01328-z

    work page doi:10.1038/s41597-022-01328-z 2022

  53. [55]

    Vision meets robotics: The KITTI dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The KITTI dataset,” Int. J. Robot. Res. , vol. 32, no. 11, pp. 1231 –1237, Sep. 2013, doi: 10.1177/0278364913491297

    doi:10.1177/0278364913491297 2013

  54. [57]

    P. K. Khani and M. J. S. Jafari Siavoshani, Proposing a numerical metric for comparing and evaluating interpretability methods of machine learning models . Sharif University of Technology, 2023. [Online]. Available: https://library.sha- rif.ir/parvan/resource/503386/. Accessed: Jun. 02, 2025

    work page 2023

  55. [2025]

    Available: https://openaccess.thecvf.com/con- tent_iccv_2015/html/Girshick_Fast_R-CNN_ICCV_2015_paper.html

    [Online]. Available: https://openaccess.thecvf.com/con- tent_iccv_2015/html/Girshick_Fast_R-CNN_ICCV_2015_paper.html

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.