Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Influence of color correction on pathology detection in Capsule Endoscopy

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Color correction makes capsule-endoscopy detectors draw bigger boxes and fire more false positives, with no consistent gain in F1, IoU, or AP50.

desk verdict Color correction shifts detection behavior in capsule endoscopy, but the paper's larger-boxes and more-FPs claims are undercut by an unreported confidence threshold. read the letter →

arxiv 2502.00076 v1 pith:D4ZNE5MR submitted 2025-01-31 cs.CV cs.AIcs.LG

Bidossessi Emmanuel Agossou , Marius Pedersen , Kiran Raja , Anuja Vats , P{aa}l Anders Floor This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords colorcorrectionwirelesscapsuleendoscopyobjectdetectionRetinanetYOLOv5boundingboxesfalsepositivesmedicalimagepreprocessing
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 tests whether correcting the colors of wireless capsule endoscopy images before training object detectors changes how well the detectors find pathologies. It builds two color-corrected versions of the SEE-AI dataset and trains Retinanet and YOLOv5 on the original and corrected versions. The central finding is that color correction changes the geometry of detections: models produce larger bounding boxes and larger overlap areas with ground truth, and false positives increase for several pathologies such as bleeding, erosion, erythema, and polyp-like lesions. Yet these changes do not translate into consistent gains in F1-score, IoU, or AP50. This matters because standard summary metrics would miss the geometric and false-alarm effects of a preprocessing choice.

What carries the argument

The machinery is the pair of color correction matrices: one derived from a standard 24-patch ColorChecker and one from a colon-specific ColorChecker with 24 gastrointestinal colors plus six gray patches, both taken from the cited color-correction study. These matrices are applied to every SEE-AI frame to produce the two color-corrected datasets. The evaluation then relies on two one-stage detectors, Retinanet with focal loss and YOLOv5, and on combined sign tests that compare, per pathology, the number of false positives and the pixel areas of predicted boxes across the three color schemes. The color corrections intensify reddish hues but reduce contrast, which the paper links to the observed changes in false positives.

What would settle it

Re-run the same experiments while sweeping the detection confidence threshold from low to high and record false-positive counts, bounding-box areas, and intersection areas for each color scheme; if the ordering of the original and two corrected versions reverses or disappears at some thresholds, the claimed effects are threshold artifacts rather than genuine consequences of color correction.

Watch

Extended reading notes

Core claim

The paper's claim, on its own terms, is that applying two color correction functions to the SEE-AI dataset significantly changes detector behavior even though headline metrics barely move. Specifically, color-corrected training data leads Retinanet and YOLOv5 to output larger bounding boxes and larger intersection areas with ground-truth boxes, and for selected pathologies the number of false positives rises; combined sign tests show these differences are statistically significant. The same corrections do not yield a consistent improvement in F1, IoU, or AP at an IoU threshold of 0.5 across the two models and twelve pathology classes. The authors conclude that color correction alone is not a reliable performance booster for WCE detection and that its main measurable effects are geometric and false-alarm shifts, with a visible loss of contrast in corrected images.

Load-bearing premise

The comparisons of false positives and bounding-box sizes assume that a single detection confidence threshold is fair across all three color schemes, but the paper never states what that threshold is.

Editorial extensions

If this is right

  • Color correction can be expected to enlarge predicted bounding boxes and increase their overlap with ground truth in WCE detection, so box-size statistics should be reported alongside F1 and AP.
  • For pathologies such as bleeding, erosion, erythema, and polyp-like lesions, corrected color tends to raise false-positive counts, so clinical screening pipelines using these corrections should expect more spurious alarms.
  • Standard metrics like F1, IoU, and AP50 can stay flat even when detection geometry changes substantially, meaning metric-only comparisons may miss preprocessing effects.
  • The loss of contrast in corrected images suggests color correction may need to be paired with contrast enhancement to be diagnostically useful.
  • Weighted aggregate scores can hide pathology-specific improvements and degradations, since some classes improve under correction while others worsen.

Reading between the lines

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

  • If the observed false-positive increase is driven by a fixed confidence threshold rather than by color itself, then reporting precision-recall curves or using confidence-calibrated thresholds could make the color-correction effect disappear or reverse; the paper does not test this.
  • A direct testable extension would measure detection confidence distributions under each color scheme: if color correction systematically raises output confidence for wrong boxes, then box size and false-positive counts are partly a calibration artifact.
  • The results suggest that evaluating preprocessing choices by final mAP alone can hide real behavioral changes; medical detection benchmarks could include box-geometry and false-alarm statistics as standard reporting.
  • Because the correction matrices were derived from a different acquisition setup, applying them to PillCam SB-3 images may not be optimal; learning a correction on SEE-AI itself could change the findings.
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 / 5 minor

Summary. The paper investigates whether two color-correction transforms (CC and CCC, taken from Watine et al. [8]) applied to the SEE-AI wireless capsule endoscopy dataset change the pathology-detection behavior of RetinaNet and YOLOv5. The authors train both detectors on the original, CCD, and CCCD versions of the dataset and compare F1, AP, and IoU, then analyze false-positive counts, bounding-box areas, and intersection areas with ground truth. The central claim is that color correction changes detection geometry (larger boxes and larger intersection areas) and increases false positives for some pathologies, while not producing consistent improvements in F1/AP/IoU.

Significance. If substantiated, the result is valuable for the WCE community and for object-detection practice generally: it shows that an apparently benign preprocessing step can shift detection geometry and false-alarm rates without moving threshold-averaged summary metrics. The paper ships code, uses two standard detectors, and evaluates on a public dataset, which supports reproducibility. The primary empirical pattern—that F1/AP/IoU are not consistently changed while false-positive counts and box statistics vary—is plausible and worth reporting, provided the operating-point ambiguity is resolved.

major comments (3)
  1. [Section 3.3 and Section 4 (Tables 4-6, Figures 7-8)] The manuscript never states the detection confidence threshold used to count a predicted box as a detection when computing false-positive counts, bounding-box areas, and intersection areas. All of these quantities are operating-point-dependent. If color correction recalibrates confidence scores, a single fixed threshold can make one color scheme appear to produce more false positives and larger boxes simply because the model emits more low-confidence detections. Please report the exact threshold (and whether it was tuned per model or color scheme), and provide a sensitivity analysis (e.g., FP counts and median box areas as functions of the confidence threshold) to show that the conclusions are not artifacts of a particular operating point.
  2. [Abstract and Section 4 (Table 5, Figures 7-8)] The claim that color correction 'makes the models generate larger bounding boxes' is not supported by the evidence as presented. Table 5 reports only pairwise p-values from a combined sign test, with no direction or effect size, so it shows that box areas differ, not that they are larger. The intersection-area claim is illustrated for only two pathologies (Polyp-Like and Bleeding), and Table 3 shows both increases and decreases in IoU across classes (e.g., erosion and bleeding decrease with both corrections). Please report the median/mean box areas and intersection areas per condition with direction and confidence intervals, and restrict the abstract claim to the pathologies actually measured.
  3. [Section 4, Table 6] The conclusion that color correction increases the number of false positives is based on raw counts without per-class statistical testing or a test of the aggregate direction. The table itself shows exceptions: for Polyp-Like with RetinaNet, R-CCCD has fewer false positives than R-OrigD (90 vs. 122), and for Bleeding with YOLOv5, R-CCD has fewer than R-OrigD (17 vs. 18). The text also states that the effect is 'dependent on the specific characteristics of the pathologies' and later generalizes. Please provide per-class sign-test results with directions and a statistical comparison of the totals, so the reader can see which classes drive the aggregate effect.
minor comments (5)
  1. [Section 4] The 'combined sign test' is invoked but never described: what statistic is combined, how images are paired across color schemes, and whether any multiple-comparison correction is applied. Please add a short methodological description.
  2. [Section 3.3] Equation (1) combines the precision and recall definitions in one line, which is unclear; consider splitting them. Also, 'AP with 50% threshold' is nonstandard notation; use AP50.
  3. [General] The detector name is spelled inconsistently ('Retinanet' vs 'RetinaNet'); please standardize.
  4. [Figure 9] Figure 9 is referenced in the text but has no caption or description in the manuscript; please add a caption and explain what is shown.
  5. [Section 3.2] Training details such as the number of epochs, learning rates, augmentations, and the exact split are not given in the paper; the GitHub repository is mentioned, but the manuscript should report these or explicitly state that they are available there.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the color-correction treatment is an external input, and detection metrics are independently measured on a public benchmark; only a minor non-load-bearing self-citation to Watine et al. is present.

full rationale

The paper's central claims are empirical measurements of detector behavior on original versus color-corrected versions of an external benchmark (SEE-AI [10]). The color-correction matrices are taken from Watine et al. [8], a prior paper by two of the present co-authors; this is a self-citation, but it is not load-bearing for the detection conclusions. The matrices are fixed external inputs derived from a color-checker setup, not fitted to the SEE-AI data or to the paper's outcome metrics. The models are retrained from scratch on each dataset, and the summary metrics (F1, AP50, IoU) are standard external criteria computed against ground-truth annotations. No equation in the paper reduces a claimed result to an input: the bounding-box area, intersection-area, and false-positive counts are observed statistics, not quantities defined by the color-correction transformation. The absence of a stated confidence threshold for counting detections in Tables 4-6 and Figures 7-8 is a potential validity limitation, because a fixed operating point can make color correction look like it changes box counts via confidence recalibration; however, this is an underdetermination or correctness issue, not circularity, since the paper does not fit a threshold and then rename it a prediction. Table 5 reports only p-values without effect direction, and the intersection-area claim is illustrated for only two pathologies, but these are reporting gaps rather than circular reductions. The self-citation to [8] therefore warrants at most a minor score.

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

The central claims rest on two pre-existing color correction matrices, an unreported confidence threshold, and the correctness of the public SEE-AI annotations. No new physical or conceptual entities are introduced.

free parameters (3)
  • CC color correction matrix = 3x3 matrix from Watine et al. [8]
    The conclusions are specific to this correction matrix, which was derived for a standard color checker, not for the SEE-AI camera.
  • CCC color correction matrix = 3x3 matrix from Watine et al. [8]
    The conclusions are specific to the colon-color-checker correction; a different correction method could give different results.
  • Detection confidence threshold = not reported
    Used to count false positives and measure bounding box areas in Section 4; if it differs across models, the FP and box-size comparisons may be confounded.
assumptions (5)
  • domain assumption The ground truth bounding boxes in the SEE-AI dataset are correct and complete.
    All metrics and FP counts are computed against these annotations; errors in annotation would bias the results.
  • domain assumption The color correction matrices [8] were applied faithfully to all images, with no additional undocumented preprocessing.
    The paper describes the corrections but does not provide the implementation or check that out-of-gamut effects are handled identically.
  • domain assumption Training and test splits maintain class balance across the 12 pathologies.
    Stated in Section 3.3; if the split is not representative, the metrics could be misleading.
  • standard math The combined sign test is a valid statistical procedure for the paired differences used.
    The test is mentioned without derivation; its assumptions (e.g., independence across images) are not verified.
  • domain assumption The models were trained to convergence with default hyperparameters.
    The paper does not report training epochs, learning rate, or early stopping; results could vary with hyperparameter choices.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Influence of color correction on pathology detection in Capsule Endoscopy." pith.science (2026). https://pith.science/paper/D4ZNE5MR

@misc{pith2026250200076,
  author       = {Pith},
  title        = {Pith review of: Influence of color correction on pathology detection in Capsule Endoscopy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4ZNE5MR}},
  note         = {Machine review of arXiv:2502.00076}
}
read the original abstract

Pathology detection in Wireless Capsule Endoscopy (WCE) using deep learning has been explored in the recent past. However, deep learning models can be influenced by the color quality of the dataset used to train them, impacting detection, segmentation and classification tasks. In this work, we evaluate the impact of color correction on pathology detection using two prominent object detection models: Retinanet and YOLOv5. We first generate two color corrected versions of a popular WCE dataset (i.e., SEE-AI dataset) using two different color correction functions. We then evaluate the performance of the Retinanet and YOLOv5 on the original and color corrected versions of the dataset. The results reveal that color correction makes the models generate larger bounding boxes and larger intersection areas with the ground truth annotations. Furthermore, color correction leads to an increased number of false positives for certain pathologies. However, these effects do not translate into a consistent improvement in performance metrics such as F1-scores, IoU, and AP50. The code is available at https://github.com/agossouema2011/WCE2024. Keywords: Wireless Capsule Endoscopy, Color correction, Retinanet, YOLOv5, Detection

Figures

Figures reproduced from arXiv: 2502.00076 by the authors.

Figure 1
Figure 1. SEE-AI Dataset Annotations per class. We can observe that classes like lymph￾follicle, erosion and polyp-like have more annotations, where as SMT, stenosis, and diverticulum have very few annotations [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Original images with their CC and CCC corrected versions. Compared to the original images, a loss of contrast can be observed with color corrected images [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Retinanet F1-scores across dif￾ferent color schemes. The F1 scores in￾crease with both color corrections (CC and CCC) for angiodysplasia, lymphang￾iectasia and vein. Only CC increases F1- score for SMT while CCC increases F1- score polyp-like. 0 0.1 0 0 0 0 0 0 0.8 0.9 angiodysplasia erosion stenosis lymphangiectasia lymph-follicle SMT polyp-like bleeding diverticulum erythema foreign-body vein YOLOv5: F1 Scores… view at source ↗
Figures from the paper (4 more)
Figure 6
Figure 6. Figure 6: YOLOv5: AP across different color schemes. The AP increase with both color corrections (CC and CCC) for lym￾phangiectasia, polyp-like and bleeding. Only CC increases AP for stenosis, while CCC increases AP for erosion and an￾giodysplasia [PITH_FULL_IMAGE:figures/full_…
Figure 7
Figure 7. Figure 7: Violin plots for Polyp-Like. With Retinanet and YOLOv5, R-CCD creates larger intersection area with the ground truth than the R-CCCD, which creates larger intersection area than the R-OrigD [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Violin plots for Bleeding. With Retinanet and YOLOv5, R-CCCD creates larger intersection area than R-CCD and R-OrigD [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Models detection results across the three different color schemes [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages

  1. [8]

    Enhancement of colour reproduction for capsule endoscopy images,

    L. Watine, P. A. Floor, M. Pedersen, P. Nussbaum, B. Ahmad, and Ø. Hovde, “Enhancement of colour reproduction for capsule endoscopy images,” in 2023 11th European Workshop on Visual Information Processing (EUVIP ). IEEE, 2023, pp. 1–6

    work page 2023

  2. [1]

    Global, regional, and national burden of 10 digestive diseases in 204 countries and territories from 1990 to 2019,

    R. Wang, Z. Li, S. Liu, and D. Zhang, “Global, regional, and national burden of 10 digestive diseases in 204 countries and territories from 1990 to 2019,” Frontiers in Public Health , vol. 11, 2023

    work page 1990

  3. [2]

    Multiple ble eding detection in wireless capsule endoscopy,

    O. Bchir, M. M. Ben Ismail, and N. AlZahrani, “Multiple ble eding detection in wireless capsule endoscopy,” Signal, Image and Video Processing , vol. 13, pp. 121– 126, 2019

    work page 2019

  4. [3]

    Patient experience of gastrointestinal end oscopy: Informing the development of the newcastle endoprem™,

    L. J. Neilson, J. Patterson, C. von Wagner, P. Hewitson, L. M. McGregor, L. Sharp, and C. J. Rees, “Patient experience of gastrointestinal end oscopy: Informing the development of the newcastle endoprem™,” Frontline Gastroenterology, vol. 11, no. 3, pp. 209–217, 2020

    work page 2020

  5. [4]

    Wireless capsule endoscopy,

    G. Iddan, G. Meron, A. Glukhovsky, and P. Swain, “Wireless capsule endoscopy,” Nature, vol. 405, no. 6785, pp. 417–417, 2000

    2000

  6. [5]

    A survey on deep learning models for wireless capsule endoscopy image analysis,

    P. Muruganantham and S. M. Balakrishnan, “A survey on deep learning models for wireless capsule endoscopy image analysis,” International Journal of Cognitive Computing in Engineering , vol. 2, pp. 83–92, 2021

    work page 2021

  7. [6]

    A survey on contempor ary computer- aided tumor, polyp, and ulcer detection methods in wireless capsule endoscopy imaging,

    T. Rahim, M. A. Usman, and S. Y. Shin, “A survey on contempor ary computer- aided tumor, polyp, and ulcer detection methods in wireless capsule endoscopy imaging,” Computerized Medical Imaging and Graphics , vol. 85, p. 101767, 2020

    work page 2020

  8. [7]

    Recent developments in wireless capsule endoscopy imaging: Compression and summarization techniques,

    B. Sushma and P. Aparna, “Recent developments in wireless capsule endoscopy imaging: Compression and summarization techniques,” Computers in Biology and Medicine, vol. 149, p. 106087, 2022

    2022

Show all 29 references
  1. [9]

    Consistency and standardization of color in medical imaging: a consensus report,

    A. Badano, C. Revie, A. Casertano, W.-C. Cheng, P. Green, T . Kimpe, E. Krupin- ski, C. Sisson, S. Skrøvseth, D. Treanor et al. , “Consistency and standardization of color in medical imaging: a consensus report,” Journal of digital imaging , vol. 28, pp. 41–52, 2015

    2015

  2. [10]

    Small bowel capsule endoscopy ex- amination and open access database with artificial intellig ence: The see-artificial intelligence project,

    A. Yokote, J. Umeno, K. Kawasaki, S. Fujioka, Y. Fuyuno, Y . Matsuno, Y. Yoshida, N. Imazu, S. Miyazono, T. Moriyama et al. , “Small bowel capsule endoscopy ex- amination and open access database with artificial intellig ence: The see-artificial intelligence project,” DEN open ,...

    2024

  3. [11]

    Unbiased teacher for semi-supervised object det ection,

    Y.-C. Liu, C.-Y. Ma, Z. He, C.-W. Kuo, K. Chen, P. Zhang, B. Wu, Z. Kira, and P. Vajda, “Unbiased teacher for semi-supervised object det ection,” arXiv preprint arXiv:2102.09480, 2021

    2021 arXiv

  4. [12]

    Instant-teach ing: An end-to-end semi-supervised object detection framework,

    Q. Zhou, C. Yu, Z. Wang, Q. Qian, and H. Li, “Instant-teach ing: An end-to-end semi-supervised object detection framework,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 4081–4090

    2021

  5. [13]

    Active teacher for semi-supervised object detectio n,

    P. Mi, J. Lin, Y. Zhou, Y. Shen, G. Luo, X. Sun, L. Cao, R. Fu, Q. Xu, and R. Ji, “Active teacher for semi-supervised object detectio n,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni tion, 2022, pp. 14 482–14 491. Influence of color correctio...

    2022

  6. [14]

    You o nly look once: Unified, real-time object detection,

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

    2016

  7. [15]

    Ssd: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. F u, and A. C. Berg, “Ssd: Single shot multibox detector,” in Computer Vision–ECCV 2016: 14th Euro- pean Conference, Amsterdam, The Netherlands, October 11–1 4, 2016, Proceedings, Part I 14 . Springer, 2016, pp. 21–37

    2016

  8. [16]

    Faster r-cnn: Towa rds real-time object de- tection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towa rds real-time object de- tection with region proposal networks,” Advances in neural information processing systems, vol. 28, 2015

    2015

  9. [17]

    Mask r-cn n,

    K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask r-cn n,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2961–2969

    2017

  10. [18]

    Gastroenterologist-level identification of small-bowe l diseases and normal variants by capsule endoscopy using a deep-learning model,

    Z. Ding, H. Shi, H. Zhang, L. Meng, M. Fan, C. Han, K. Zhang, F. Ming, X. Xie, H. Liu et al. , “Gastroenterologist-level identification of small-bowe l diseases and normal variants by capsule endoscopy using a deep-learning model,” Gastroen- terology, vol. 157, no. 4, pp. 1044–...

    2019

  11. [19]

    Automatic detection of blood content in capsule endoscopy images based on a deep convolutional neural netwo rk,

    T. Aoki, A. Yamada, Y. Kato, H. Saito, A. Tsuboi, A. Nakada , R. Niikura, M. Fu- jishiro, S. Oka, S. Ishihara et al. , “Automatic detection of blood content in capsule endoscopy images based on a deep convolutional neural netwo rk,” Journal of gas- troenterology and hepatology...

    2020

  12. [20]

    Automatic detection of various abnormalities in ca psule endoscopy videos by a deep learning-based system: a multicenter study,

    ——, “Automatic detection of various abnormalities in ca psule endoscopy videos by a deep learning-based system: a multicenter study,” Gastrointestinal Endoscopy, vol. 93, no. 1, pp. 165–173, 2021

    2021

  13. [21]

    Enhancing dermoscop y images to improve melanoma detection,

    O. Cherepkova and J. Y. Hardeberg, “Enhancing dermoscop y images to improve melanoma detection,” in 2018 Colour and Visual Computing Symposium (CVCS) . IEEE, 2018, pp. 1–6

    2018

  14. [22]

    Fo cal loss for dense object detection,

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Fo cal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2980–2988

    2017

  15. [23]

    M. logo. Pillcam™ sb 3 capsule endoscopy system. [Online ]. A vailable: https://www.medtronic.com/covidien/en-us/products/capsule-endoscopy/pillcam-sb3-system.html

  16. [24]

    Rethinking dice loss for medical image segmentation,

    R. Zhao, B. Qian, X. Zhang, Y. Li, R. Wei, Y. Liu, and Y. Pan, “Rethinking dice loss for medical image segmentation,” in 2020 IEEE International Conference on Data Mining (ICDM) . IEEE, 2020, pp. 851–860

    2020

  17. [25]

    Ratio-and-scale-aware yolo fo r pedestrian detection,

    W.-Y. Hsu and W.-Y. Lin, “Ratio-and-scale-aware yolo fo r pedestrian detection,” IEEE transactions on image processing , vol. 30, pp. 934–947, 2020

    2020

  18. [26]

    Top ten challenges in object detec- tion every data scientist should know

    S. Mukherjee, “Top ten challenges in object detec- tion every data scientist should know.” [Online]. A vailabl e: https://analyticsindiamag.com/top-ten-challenges-in-object-detection-every-data-scientist-should-know/

  19. [27]

    A. Anwar. What is average precision in object detection & lo- calization algorithms and how to calculate it? [Online]. A v ailable: https://towardsdatascience.com/what-is-average-precision-in-object-detection-localization-algorithms-a nd-how-to-

  20. [28]

    Per formance analysis of deep learning-based object detection algorithms on coco be nchmark: a comparative study,

    J. Tian, Q. Jin, Y. Wang, J. Yang, S. Zhang, and D. Sun, “Per formance analysis of deep learning-based object detection algorithms on coco be nchmark: a comparative study,” Journal of Engineering and Applied Science , vol. 71, no. 1, p. 76, 2024

    2024

  21. [29]

    Intersection over union (iou) in ob- ject detection & segmentation

    Kukil. Intersection over union (iou) in ob- ject detection & segmentation. [Online]. A vailable: https://learnopencv.com/intersection-over-union-iou-in-object-detection-and-segmentation/ Ϭ ϱ ϭϬ ϭϱ ϮϬ Ϯϱ ϯϬ ϯϱ ϰϬ ϰϱ ϱϬ Ϭ ϱ ϭϬ ϭϱ ϮϬ Ϯϱ ϯϬ ĂƚĂ ĂƚĂ

Pith tools

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