pith. sign in

arxiv: 2605.26682 · v1 · pith:T3M6VVOQnew · submitted 2026-05-26 · 💻 cs.RO · cs.CV

SteelDS: A High-Resolution Video Dataset of E40 Steel Scrap for Object Detection and Instance Segmentation

Melanie Neubauer , Christian Rauch , Gerald Koinig , Alexia Tischberger-Aldrian , Roland Pomberger , Elmar Rueckert This is my paper

Pith reviewed 2026-06-29 17:09 UTC · model grok-4.3

classification 💻 cs.RO cs.CV
keywords SteelDSsteel scrap datasetobject detectioninstance segmentationcopper impuritiesconveyor beltrecycling automationmaterial classification
0
0 comments X

The pith

SteelDS supplies 24,297 annotated video frames of E40 steel and copper scrap to benchmark automated impurity detection on conveyor belts.

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

The paper releases SteelDS, a collection of high-resolution video sequences showing shredded E40 steel mixed with copper objects moving on a conveyor. The sequences are recorded under controlled conditions that replicate the post-magnetic sorting stage of industrial recycling, where copper contaminants must still be removed by hand. Each frame carries pixel-level segmentation masks and material labels for steel and copper items of varying sizes, with deliberate changes in object spacing and density. A reader would care because the dataset supplies the training and test material needed to build machine-vision systems that could replace or assist that manual step. If the dataset works as intended, algorithms trained on it can be evaluated directly on their ability to locate copper inside realistic scrap streams.

Core claim

The authors present SteelDS as a benchmark dataset whose 24,297 labeled frames across five subsets, containing 396 steel and 101 copper objects, enable quantitative evaluation of object detection, instance segmentation, and material classification models for the specific task of identifying copper impurities in heterogeneous E40 steel scrap on a conveyor belt.

What carries the argument

The SteelDS dataset itself, consisting of pixel-wise segmentation masks and material-class labels for video frames recorded under controlled variations of object spacing and density.

If this is right

  • Algorithms can be trained and scored on the joint tasks of localizing objects and assigning steel versus copper labels.
  • Performance can be measured across different object densities and spacings that the dataset explicitly varies.
  • The pixel-level masks allow direct comparison of instance segmentation methods against the same ground truth.
  • The five subsets provide separate training, validation, and test partitions for reproducible benchmarking.

Where Pith is reading between the lines

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

  • Robotic systems could use models trained on this data to trigger selective removal of copper pieces without stopping the belt.
  • The same annotation format could be reused for other scrap grades or additional contaminant types once similar video is collected.
  • Accuracy on SteelDS may serve as a quick filter before more expensive real-plant trials of any new sorting algorithm.

Load-bearing premise

The laboratory recordings with their chosen spacing and density variations are representative of the real industrial post-magnetic sorting stage.

What would settle it

Models that reach high accuracy on SteelDS yet show low accuracy when tested on video recorded directly from an operating industrial sorting line after magnetic separation.

read the original abstract

This dataset provides high-resolution, annotated video sequences of shredded E40-grade steel and copper scrap on a conveyor belt. Captured in a controlled laboratory environment, the data reflects the industrial post-magnetic sorting stage, where manual intervention is typically required to remove copper contaminants. The dataset comprises 24,297 labeled frames across five subsets, featuring 396 steel and 101 copper objects categorized by size. It supports the development of machine learning models for material classification, object detection, and instance segmentation. Variations in object spacing and density are included to simulate realistic industrial sorting conditions. Ground truth annotations include pixel-wise segmentation masks and material classes. This dataset serves as a benchmark for evaluating automated sorting algorithms aiming to identify copper impurities within complex, heterogeneous steel scrap streams.

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 presents SteelDS, a dataset of 24,297 high-resolution annotated video frames of shredded E40 steel scrap mixed with copper objects on a conveyor belt. Captured in a controlled laboratory setting to approximate the post-magnetic sorting stage, the data includes pixel-wise segmentation masks and material class labels (steel vs. copper), with objects categorized by size and variations in spacing/density across five subsets. The central claim is that SteelDS serves as a benchmark for developing and evaluating ML models for object detection, instance segmentation, and material classification to identify copper impurities in heterogeneous steel scrap streams.

Significance. If the annotations prove reliable and the scenes sufficiently representative, the dataset would address a gap in publicly available, application-specific data for industrial recycling automation. The video format, high resolution, and controlled variations in object density provide a foundation for training robust detection models that could reduce manual sorting needs; the explicit scoping to the post-magnetic stage makes the intended use case clear.

major comments (2)
  1. [Dataset description / abstract] Dataset description (abstract and § on data collection): no information is supplied on the annotation process, annotator qualifications, tools employed, inter-annotator agreement, or quality-control procedures. Without these details the ground-truth masks and class labels cannot be verified, directly undermining the claim that the dataset functions as a reliable benchmark.
  2. [Dataset description / abstract] Dataset description (abstract): the statement that the laboratory captures "reflect the industrial post-magnetic sorting stage" is unsupported by any quantitative validation, sensor comparison, or material-composition statistics against real plant data. This assumption is load-bearing for the benchmark claim yet remains untested.
minor comments (1)
  1. [Dataset description] The five subsets are mentioned but their distinguishing characteristics (e.g., exact density ranges or camera angles) are not tabulated, reducing reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for these constructive comments on the dataset description. Both points identify important omissions that weaken the benchmark claim, and we will revise the manuscript to address them directly.

read point-by-point responses

  1. Referee: [Dataset description / abstract] Dataset description (abstract and § on data collection): no information is supplied on the annotation process, annotator qualifications, tools employed, inter-annotator agreement, or quality-control procedures. Without these details the ground-truth masks and class labels cannot be verified, directly undermining the claim that the dataset functions as a reliable benchmark.

    Authors: We agree the annotation details are missing and essential. The revised manuscript will add a dedicated subsection describing the annotation workflow, tools, annotator qualifications, inter-annotator agreement metrics, and quality-control steps. revision: yes

  2. Referee: [Dataset description / abstract] Dataset description (abstract): the statement that the laboratory captures "reflect the industrial post-magnetic sorting stage" is unsupported by any quantitative validation, sensor comparison, or material-composition statistics against real plant data. This assumption is load-bearing for the benchmark claim yet remains untested.

    Authors: The comment is correct; no quantitative validation is supplied. We will revise the abstract and data-collection section to qualify the claim, describing the laboratory setup as an approximation of the post-magnetic stage based on process similarity while explicitly noting the lack of direct plant-data comparisons. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

This is a dataset release paper with no derivations, equations, fitted parameters, predictions, or load-bearing self-citations. The abstract and description explicitly scope the work to controlled laboratory capture of E40 steel scrap with stated variations in spacing/density, presented as a benchmark without any internal chain that reduces a claimed result to its own inputs by construction. No patterns from the enumerated circularity kinds apply.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Dataset release paper; contains no mathematical derivations, fitted parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5679 in / 955 out tokens · 46269 ms · 2026-06-29T17:09:21.397769+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

26 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    & Beyerer, J

    Maier, G., Gruna, R., Längle, T. & Beyerer, J. A survey of the state of the art in sensor-based sorting technology and research.IEEE access12, 6473–6493 (2024)

    2024

  2. [2]

    Han, S. D.et al.Toward fully automated metal recycling using computer vision and non-prehensile manipulation.2021 IEEE 17th International Conference on Automation Science and Engineering (CASE)891–898 (2021)

    2021

  3. [3]

    K., Singh, R

    Sarswat, P. K., Singh, R. S. & Pathapati, S. V. S. H. Real time electronic-waste classification algorithms using the computer vision based on convolutional neu- ral network (cnn): Enhanced environmental incentives.Resources, Conservation and Recycling207, 107651 (2024). URL https://www.sciencedirect.com/science/ article/pii/S0921344924002453

    2024

  4. [4]

    Proença, P. F. & Simões, P. Taco: Trash annotations in context for litter detection.arXiv preprint arXiv:2003.06975(2020)

    work page arXiv 2003

  5. [5]

    & Thung, G

    Yang, M. & Thung, G. Classification of trash for recyclability status.CS229 project report2016, 3 (2016). URL https://cs229.stanford.edu/proj2016/report/ ThungYang-ClassificationOfTrashForRecyclabilityStatus-report.pdf

    2016

  6. [6]

    Bashkirova, D.et al.Zerowaste dataset: Towards deformable object segmenta- tion in cluttered scenes.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)21115–21125 (2022). 19

    2022

  7. [7]

    URL https://doi.org/10.1038/ s41597-025-05243-x

    Sirimewan, D., Dayarathna, S., Raman, S.et al.A benchmark dataset for class-wise segmentation of construction and demolition waste in cluttered envi- ronments.Scientific Data12, 885 (2025). URL https://doi.org/10.1038/ s41597-025-05243-x

    2025

  8. [8]

    & Makarov, I

    Storonkin, D., Dziub, I., Golyadkin, M. & Makarov, I. From images to decisions: Assistive computer vision for non-metallic content estimation in scrap metal. arXiv preprint arXiv:2602.07062(2026)

    work page arXiv 2026

  9. [9]

    & Hua, L

    Chen, S., Hu, Z., Wang, C., Pang, Q. & Hua, L. Research on the process of small sample non-ferrous metal recognition and separation based on deep learning. Waste Management126, 266–273 (2021). URL https://www.sciencedirect.com/ science/article/pii/S0956053X21001616

    2021

  10. [10]

    URL https://www.sciencedirect.com/science/article/pii/S0956053X24005439

    Koinig, G.et al.Deep learning approaches for classification of copper-containing metal scrap in recycling processes.Waste Management190, 520–530 (2024). URL https://www.sciencedirect.com/science/article/pii/S0956053X24005439

    2024

  11. [11]

    Koinig, G.et al.Cnn-based copper reduction in shredded scrap for enhanced elec- tric arc furnace steelmaking.OCM 2025-7th International Conference on Optical Characterization of Materials, March 26th–27th, 2025, Karlsruhe, Germany: Conference Proceedings329 (2025)

    2025

  12. [12]

    & Eder, P

    Muchová, L. & Eder, P. End-of-waste criteria for iron and steel scrap: Technical proposals. Tech. Rep., Joint Research Centre (JRC), Institute for Prospective Technological Studies, Luxembourg (2010). Publications Office of the European Union

    2010

  13. [13]

    Aboussouan, L.et al.Steel scrap fragmentation by shredders.Powder Technology 105, 288–294 (1999)

    1999

  14. [14]

    & Glaser, B

    Schäfer, M., Faltings, U. & Glaser, B. Does-a multimodal dataset for supervised and unsupervised analysis of steel scrap.Scientific data10, 780 (2023)

    2023

  15. [15]

    & Wiczyński, G

    Jurtsch, T., Moryson, J. & Wiczyński, G. Machine vision-based detection of forbidden elements in the high-speed automatic scrap sorting line.Waste Management189, 243–253 (2024)

    2024

  16. [16]

    & Rückert, E

    Neubauer, M. & Rückert, E. Semi-autonomous fast object segmentation and tracking tool for industrial applications.2024 21st International Conference on Ubiquitous Robots (UR)77–83 (2024)

    2024

  17. [17]

    URL https://doi.org/10.5281/zenodo.20271102

    Neubauer, M.et al.SteelDS: A High-Resolution Video Dataset of E40 Steel Scrap for Object Detection and Instance Segmentation.Zenodo(2026). URL https://doi.org/10.5281/zenodo.20271102. 20

    work page doi:10.5281/zenodo.20271102 2026

  18. [18]

    & Farhadi, A

    Redmon, J., Divvala, S., Girshick, R. & Farhadi, A. You only look once: Uni- fied, real-time object detection.Proceedings of the IEEE conference on computer vision and pattern recognition779–788 (2016)

    2016

  19. [19]

    & Qiu, J

    Jocher, G., Chaurasia, A. & Qiu, J. Ultralytics yolov8 (2023). URL https: //github.com/ultralytics/ultralytics. Version 8.0.0

    2023

  20. [20]

    & Qiu, J

    Jocher, G. & Qiu, J. Ultralytics yolo11 (2024). URL https://github.com/ ultralytics/ultralytics. Version 11.0.0

    2024

  21. [21]

    YOLOv12: Attention-Centric Real-Time Object Detectors

    Tian, Y., Ye, Q. & Doermann, D. Yolo12: Attention-centric real-time object detectors.arXiv preprint arXiv:2502.12524(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    & Qiu, J

    Jocher, G. & Qiu, J. Ultralytics yolo26 (2026). URL https://github.com/ ultralytics/ultralytics. Version 26.0.0

    2026

  23. [23]

    & Girshick, R

    He, K., Gkioxari, G., Dollár, P. & Girshick, R. Mask r-cnn.Proceedings of the IEEE international conference on computer vision2961–2969 (2017)

    2017

  24. [24]

    Lin, T.-Y.et al.Microsoft coco: Common objects in context.European conference on computer vision740–755 (2014)

    2014

  25. [25]

    PASTA: Vision Transformer Patch Aggregation for Weakly Supervised Target and Anomaly Segmentation

    Neubauer, M., Rueckert, E. & Rauch, C. Pasta: Vision transformer patch aggre- gation for weakly supervised target and anomaly segmentation (2026). URL https://arxiv.org/abs/2604.09701. arXiv:2604.09701

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  26. [26]

    & Rueckert, E

    Neubauer, M., Özdenizci, O., Piater, J. & Rueckert, E. Sparsifying instance seg- mentationmodelsforefficientvision-basedindustrialrecycling.Machine Learning and Knowledge Discovery in Databases. Applied Data Science Track and Demo Track (ECML PKDD)(2025). 21 Subset a1 Subset a2 Subset a3 Ground T ruth YOLOv8n YOLO11n YOLO12n YOLO26n Mask R-CNN Fig. 8: Vis...

    2025