pith. sign in

arxiv: 2605.26884 · v1 · pith:6VBPJ56Wnew · submitted 2026-05-26 · 💻 cs.CV

Small Object Detection in Industrial Recycling: A New Dataset and YOLO Performance Evaluation

Oussama Messai , Abbass Zein-Eddine , Abdelouahid Bentamou , Mickael Picq , Nicolas Duquesne , St\'ephane Puydarrieux , Yann Gavet This is my paper

Pith reviewed 2026-06-29 18:05 UTC · model grok-4.3

classification 💻 cs.CV
keywords small object detectionindustrial recyclingYOLOobject detection datasetdata augmentationanomaly detectiondense overlapping objects
0
0 comments X

The pith

A new dataset of over 10k images and 120k instances enables direct comparison of YOLO detectors for small dense objects in recycling.

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

The paper introduces a specialized dataset capturing small, dense, and overlapping objects from industrial recycling streams. It systematically evaluates multiple deep learning object detectors with emphasis on YOLO architectures, measuring detection accuracy, processing speed, and suitability for real-time industrial use. The evaluation also measures gains from data augmentation and synthetic image generation. Anomaly detection components are shown to maintain performance despite changes in image resolution or zoom. The work identifies top-performing systems for the recycling setting and outlines directions for further improvement.

Core claim

The paper establishes a benchmark dataset of more than 10,000 images and 120,000 instances and uses it to compare supervised deep learning detectors, particularly YOLO variants, on accuracy, computational efficiency, and robustness for small-object detection in recycling. It reports that data augmentation and synthetic images improve results, identifies the most reliable current systems, and demonstrates an anomaly detection approach that remains stable across resolution and zoom variations while extending the scope to length measurement.

What carries the argument

The new recycling-specific dataset paired with a side-by-side evaluation of YOLO and other deep learning detectors on accuracy, speed, and augmentation effects.

If this is right

  • Data augmentation and synthetic images measurably raise detection accuracy on the provided dataset.
  • Specific YOLO configurations deliver the best trade-off between precision and runtime for dense small-object scenes.
  • The anomaly detection component continues to function reliably when image resolution or zoom changes.
  • Object detection outputs can be combined with length measurement within the same recycling workflow.

Where Pith is reading between the lines

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

  • The dataset and evaluation protocol could be replicated for other manufacturing inspection tasks that involve small dense items.
  • Cross-plant testing would reveal whether the current top-ranked detectors generalize beyond the original data source.
  • Integrating the detectors with on-line length measurement hardware could create an end-to-end quality-control pipeline.

Load-bearing premise

The collected images represent the full range of object sizes, densities, overlaps, and lighting conditions encountered in operating recycling facilities.

What would settle it

Run the same set of detectors on a fresh collection of images sampled from a different recycling plant and verify whether the accuracy and efficiency rankings stay the same.

Figures

Figures reproduced from arXiv: 2605.26884 by Abbass Zein-Eddine, Abdelouahid Bentamou, Mickael Picq, Nicolas Duquesne, Oussama Messai, St\'ephane Puydarrieux, Yann Gavet.

Figure 1
Figure 1. Figure 1: Example of small, dense and overlapped objects in an industrial [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual examples of the main object classes for detection in the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Heatmap showing the distribution of bounding box annotations across [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Examples of labeled real and synthetic images. Top: real images, [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: F1-confidence curve on different image resolution using UDD v1 and [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Detection performance exhibited a steady increase as the number [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Canny edges before and after Histogram Equalization (HE). Left: [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Confusion matrix for model trained only on real data. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Confusion matrix for model pre-trained on synthetic data and fine [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Identify and quantify long metallic hulls within a collection of [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Outliers example in the bounding box diagonal distribution. Green [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Anomaly detection with IQR over bounding box diagonals. The left [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
read the original abstract

In this paper, we address the problem of detecting small, dense, and overlapping objects, a major challenge in computer vision. Our focus is on reviewing proposed methods based on deep learning supervised approaches. We provide a detailed comparison of these systems on a new dataset of more than 10k images and 120k instances, highlighting their performance, accuracy, and computational efficiency in the industrial recycling process use case. Through this comparative analysis, we identify the most reliable systems currently available and the specific challenges they are designed to tackle. Furthermore, we explore the benefits of data augmentation and synthetic images. Based on our analysis, we also propose potential future directions and innovative solutions that could enhance the effectiveness of small, dense and overlapped object detection systems. The scope of our investigations encompasses object detection, length measurement, and anomaly detection within the context of the recycling process. The anomaly detection strategy is robust against variations in image resolution and zoom levels, ensuring reliable performance in industrial applications. The repository of the proposed dataset, methods and evaluation codes can be found at: https://github.com/o-messai/SDOOD

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.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a new dataset of more than 10k images and 120k instances focused on small, dense, and overlapping objects in industrial recycling. It performs a comparative evaluation of deep-learning detectors (with emphasis on YOLO variants) for performance, accuracy, and computational efficiency, investigates data augmentation and synthetic images, and addresses length measurement plus anomaly detection claimed to be robust against resolution and zoom variations. Future directions are proposed and code/dataset are released via GitHub.

Significance. If the dataset is shown to be representative of real recycling streams, the benchmark could usefully inform model selection for industrial small-object tasks where efficiency matters. Public release of data and code supports reproducibility. The anomaly-detection robustness claim, if backed by quantitative cross-condition tests, would add practical value.

major comments (2)
  1. [Abstract and Dataset section] Abstract and Dataset section: the central claim that the comparison reliably identifies the best systems for the industrial recycling use case requires the dataset to be representative of varied real streams (plants, lighting, densities, zoom). No sampling strategy, acquisition protocol, number of source facilities, camera/setup details, or cross-plant/domain-shift validation is described.
  2. [Abstract] Abstract: the text states that a detailed comparison of performance, accuracy, and efficiency was performed yet supplies no quantitative results, tables, error bars, or exclusion criteria, so the soundness of the ranking claims cannot be verified from the provided text.
minor comments (1)
  1. [Title and Abstract] The title specifies YOLO evaluation while the abstract refers to broader deep-learning methods; clarify the exact set of detectors compared.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below and commit to revisions that strengthen the clarity and verifiability of our claims without altering the core contributions.

read point-by-point responses

  1. Referee: [Abstract and Dataset section] Abstract and Dataset section: the central claim that the comparison reliably identifies the best systems for the industrial recycling use case requires the dataset to be representative of varied real streams (plants, lighting, densities, zoom). No sampling strategy, acquisition protocol, number of source facilities, camera/setup details, or cross-plant/domain-shift validation is described.

    Authors: We agree that explicit documentation of acquisition details is required to support claims of representativeness. In the revised manuscript we will add a dedicated 'Data Acquisition' subsection that specifies the sampling strategy, acquisition protocol, number of source facilities, camera models, lighting conditions, and zoom/density variations covered. We will also report any available cross-facility consistency checks; where full domain-shift experiments were not performed we will explicitly note this as a limitation and outline it as future work. revision: yes

  2. Referee: [Abstract] Abstract: the text states that a detailed comparison of performance, accuracy, and efficiency was performed yet supplies no quantitative results, tables, error bars, or exclusion criteria, so the soundness of the ranking claims cannot be verified from the provided text.

    Authors: We accept that the abstract must contain concise quantitative support. We will revise the abstract to include the principal numerical outcomes (e.g., mAP ranges and inference speeds for the leading YOLO variants, augmentation gains, and anomaly-detection metrics) together with a brief statement of the evaluation protocol and exclusion criteria used. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical dataset and benchmark paper

full rationale

The paper introduces a new dataset (>10k images, 120k instances) and reports empirical performance numbers for existing YOLO-family detectors on held-out test splits. No derivations, first-principles results, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The central claim is a straightforward comparison of off-the-shelf models; the evaluation protocol does not reduce to its own inputs by construction. Dataset representativeness is an external validity question, not a circularity issue.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical model or derivation; empirical benchmark paper with no free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5757 in / 982 out tokens · 24212 ms · 2026-06-29T18:05:05.658436+00:00 · methodology

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

44 extracted references · 6 canonical work pages · 3 internal anchors

  1. [1]

    Object detection with deep learning: A review,

    Z.-Q. Zhao, P. Zheng, S.-t. Xu,et al., “Object detection with deep learning: A review,”IEEE transactions on neural networks and learning systems30(11), 3212–3232 (2019)

    2019

  2. [2]

    Small-object detection in remote sensing images with end-to-end edge-enhanced gan and object detector network,

    J. Rabbi, N. Ray, M. Schubert,et al., “Small-object detection in remote sensing images with end-to-end edge-enhanced gan and object detector network,”Remote Sensing12(9), 1432 (2020)

    2020

  3. [3]

    Enhancing image quality and anomaly detection for small and dense industrial objects in nuclear recycling,

    O. Messai, A. Zein-Eddine, A. Bentamou,et al., “Enhancing image quality and anomaly detection for small and dense industrial objects in nuclear recycling,” inSeventeenth International Conference on Quality Control by Artificial Vision, T. Shimizu, T. Akashi, J. Sato,et al., Eds., 13737, 1373705, International Society for Optics and Photonics, SPIE (2025)

    2025

  4. [4]

    Application of deep learning for object detection,

    A. R. Pathak, M. Pandey, and S. Rautaray, “Application of deep learning for object detection,”Procedia computer science132, 1706– 1717 (2018)

    2018

  5. [5]

    Object detection in 20 years: A survey,

    Z. Zou, K. Chen, Z. Shi,et al., “Object detection in 20 years: A survey,” Proceedings of the IEEE111(3), 257–276 (2023)

    2023

  6. [6]

    Google earth engine for geo-big data applications: A meta-analysis and systematic review,

    H. Tamiminia, B. Salehi, M. Mahdianpari,et al., “Google earth engine for geo-big data applications: A meta-analysis and systematic review,” ISPRS journal of photogrammetry and remote sensing164, 152–170 (2020)

    2020

  7. [7]

    Roboflow (version 1.0)[soft- ware],

    B. Dwyer, J. Nelson, J. Solawetz,et al., “Roboflow (version 1.0)[soft- ware],”URL: https://roboflow. com. computer vision(2022)

    2022

  8. [8]

    Generation of synthetic digital image corre- lation images using the open-source blender software,

    D. Rohe and E. Jones, “Generation of synthetic digital image corre- lation images using the open-source blender software,”Experimental Techniques46(4), 615–631 (2022)

    2022

  9. [9]

    Improved object recognition results using sift and orb feature detector,

    S. Gupta, M. Kumar, and A. Garg, “Improved object recognition results using sift and orb feature detector,”Multimedia Tools and Applications 78(23), 34157–34171 (2019)

    2019

  10. [10]

    A geometric-based method for recognizing overlapping polygonal-shaped and semi-transparent par- ticles in gray tone images,

    O. S. Ahmad, J. Debayle, and J.-C. Pinoli, “A geometric-based method for recognizing overlapping polygonal-shaped and semi-transparent par- ticles in gray tone images,”Pattern Recognition Letters32(15), 2068– 2079 (2011)

    2068

  11. [11]

    Fusing points and lines for high perfor- mance tracking,

    E. Rosten and T. Drummond, “Fusing points and lines for high perfor- mance tracking,” inTenth IEEE International Conference on Computer Vision (ICCV’05) Volume 1,2, 1508–1515, Ieee (2005)

    2005

  12. [12]

    A computational approach to edge detection,

    J. Canny, “A computational approach to edge detection,”IEEE Transac- tions on pattern analysis and machine intelligence(6), 679–698 (1986)

    1986

  13. [13]

    A multiscale method for shape recognition of overlapping elliptical particles,

    M. De Langlard, H. Al Saddik, F. Lamadie,et al., “A multiscale method for shape recognition of overlapping elliptical particles,” in2016 23rd International Conference on Pattern Recognition (ICPR), 692–697, IEEE (2016)

    2016

  14. [14]

    Recognition of highly overlapping ellipse-like bubble images,

    M. Honkanen, P. Saarenrinne, T. Stoor,et al., “Recognition of highly overlapping ellipse-like bubble images,”Measurement Science and Tech- nology16(9), 1760 (2005)

    2005

  15. [15]

    An efficiency improved recognition algorithm for highly overlapping ellipses: Appli- cation to dense bubbly flows,

    M. De Langlard, H. Al-Saddik, S. Charton,et al., “An efficiency improved recognition algorithm for highly overlapping ellipses: Appli- cation to dense bubbly flows,”Pattern Recognition Letters101, 88–95 (2018)

    2018

  16. [16]

    Fast R-CNN

    R. Girshick, “Fast r-cnn,”arXiv preprint arXiv:1504.08083(2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  17. [17]

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

    S. Ren, K. He, R. Girshick,et al., “Faster r-cnn: Towards real-time object detection with region proposal networks,”IEEE transactions on pattern analysis and machine intelligence39(6), 1137–1149 (2016)

    2016

  18. [18]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick,et al., “You only look once: Unified, real-time object detection,” inProceedings of the IEEE conference on computer vision and pattern recognition, 779–788 (2016). 12

    2016

  19. [19]

    Yolo9000: better, faster, stronger,

    J. Redmon and A. Farhadi, “Yolo9000: better, faster, stronger,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 7263–7271 (2017)

    2017

  20. [20]

    YOLOv3: An Incremental Improvement

    J. Redmon, “Yolov3: An incremental improvement,”arXiv preprint arXiv:1804.02767(2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  21. [21]

    YOLOv4: Optimal Speed and Accuracy of Object Detection

    A. Bochkovskiy, C.-Y . Wang, and H.-Y . M. Liao, “Yolov4: Op- timal speed and accuracy of object detection,”arXiv preprint arXiv:2004.10934(2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2004

  22. [22]

    ultralytics/yolov5: v7. 0- yolov5 sota realtime instance segmentation,

    G. Jocher, A. Chaurasia, A. Stoken,et al., “ultralytics/yolov5: v7. 0- yolov5 sota realtime instance segmentation,”Zenodo(2022)

    2022

  23. [23]

    arXiv preprint arXiv:2209.02976 , year=

    C. Li, L. Li, H. Jiang,et al., “Yolov6: A single-stage object detection framework for industrial applications,”arXiv preprint arXiv:2209.02976 (2022)

    work page arXiv 2022

  24. [24]

    Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 7464–7475 (2023)

    2023

  25. [25]

    A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas,

    J. Terven, D.-M. C ´ordova-Esparza, and J.-A. Romero-Gonz ´alez, “A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas,”Machine Learning and Knowledge Extraction5(4), 1680–1716 (2023)

    2023

  26. [26]

    arXiv preprint arXiv:2402.13616 , year=

    C.-Y . Wang, I.-H. Yeh, and H.-Y . M. Liao, “Yolov9: Learning what you want to learn using programmable gradient information,”arXiv preprint arXiv:2402.13616(2024)

    work page arXiv 2024

  27. [27]

    arXiv preprint arXiv:2405.14458 , year=

    A. Wang, H. Chen, L. Liu,et al., “Yolov10: Real-time end-to-end object detection,”arXiv preprint arXiv:2405.14458(2024)

    work page arXiv 2024

  28. [28]

    Ultralytics yolo11,

    G. Jocher and J. Qiu, “Ultralytics yolo11,” https://github.com/ultralytics/ultralytics(2024)

    2024

  29. [29]

    Dc-yolov8: small-size object detection algorithm based on camera sensor,

    H. Lou, X. Duan, J. Guo,et al., “Dc-yolov8: small-size object detection algorithm based on camera sensor,”Electronics12(10), 2323 (2023)

    2023

  30. [30]

    An efficient and accurate quality inspection model for steel scraps based on dense small-target detection,

    P. Xiao, C. Wang, L. Zhu,et al., “An efficient and accurate quality inspection model for steel scraps based on dense small-target detection,” Processes12(8), 1700 (2024)

    2024

  31. [31]

    Enhancing object detection in dense images: Adjustable non-maximum suppression for single-class detection,

    K. Noh, S. K. Hong, S. Makonin,et al., “Enhancing object detection in dense images: Adjustable non-maximum suppression for single-class detection,”IEEE Access(2024)

    2024

  32. [32]

    A global-local self-adaptive network for drone-view object detection,

    S. Deng, S. Li, K. Xie,et al., “A global-local self-adaptive network for drone-view object detection,”IEEE Transactions on Image Processing 30, 1556–1569 (2020)

    2020

  33. [33]

    Delving into cluttered prohibited item detection for security inspection system,

    H. Wang, T. Jia, B. Ma,et al., “Delving into cluttered prohibited item detection for security inspection system,”IEEE Transactions on Industrial Informatics(2024)

    2024

  34. [34]

    Fast and accurate convolution neural network for detecting manufacturing data,

    Y . Djenouri, G. Srivastava, and J. C.-W. Lin, “Fast and accurate convolution neural network for detecting manufacturing data,”IEEE Transactions on Industrial Informatics17(4), 2947–2955 (2020)

    2020

  35. [35]

    Deep learning optimization method for counting overlapping rice seeds,

    J. Sun, Y . Zhang, X. Zhu,et al., “Deep learning optimization method for counting overlapping rice seeds,”Journal of Food Process Engineering 44(9), e13787 (2021)

    2021

  36. [36]

    Small-object detection based on yolo and dense block via image super-resolution,

    Z.-Z. Wang, K. Xie, X.-Y . Zhang,et al., “Small-object detection based on yolo and dense block via image super-resolution,”IEEE Access9, 56416–56429 (2021)

    2021

  37. [37]

    Overlapped pedestrian detection based on yolov5 in crowded scenes,

    W. Guo, N. Shen, and T. Zhang, “Overlapped pedestrian detection based on yolov5 in crowded scenes,” in2022 3rd International Conference on Computer Vision, Image and Deep Learning & International Conference on Computer Engineering and Applications (CVIDL & ICCEA), 412– 416, IEEE (2022)

    2022

  38. [38]

    End-to-end people detection in crowded scenes,

    R. Stewart, M. Andriluka, and A. Y . Ng, “End-to-end people detection in crowded scenes,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2325–2333 (2016)

    2016

  39. [39]

    Long Short-Term Memory,

    S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,”Neural Computation9, 1735–1780 (1997)

    1997

  40. [40]

    People-tracking-by-detection and people-detection-by-tracking,

    M. Andriluka, S. Roth, and B. Schiele, “People-tracking-by-detection and people-detection-by-tracking,” in2008 IEEE Conference on com- puter vision and pattern recognition, 1–8, IEEE (2008)

    2008

  41. [41]

    Nms by representative region: Towards crowded pedestrian detection by proposal pairing,

    X. Huang, Z. Ge, Z. Jie,et al., “Nms by representative region: Towards crowded pedestrian detection by proposal pairing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10750–10759 (2020)

    2020

  42. [42]

    Detrs beat yolos on real-time object detection,

    Y . Zhao, W. Lv, S. Xu,et al., “Detrs beat yolos on real-time object detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16965–16974 (2024)

    2024

  43. [43]

    Adaptive histogram equalization and its variations,

    S. M. Pizer, E. P. Amburn, J. D. Austin,et al., “Adaptive histogram equalization and its variations,”Computer vision, graphics, and image processing39(3), 355–368 (1987)

    1987

  44. [44]

    Detection of outliers using interquartile range technique from intrusion dataset,

    H. Vinutha, B. Poornima, and B. Sagar, “Detection of outliers using interquartile range technique from intrusion dataset,” inInformation and decision sciences: Proceedings of the 6th international conference on ficta, 511–518, Springer (2018)

    2018