Pith. sign in

REVIEW 4 major objections 6 minor 19 references

Image-Based Method For Measuring And Classification Of Iron Ore Pellets Using Star-Convex Polygons

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A medical cell-segmentation method, StarDist, reaches 0.962 IoU on iron ore pellets and links mask diameters to lab sizes within about 5 mm.

desk verdict Plausible StarDist transfer to industrial pellet inspection with promising numbers, but the unvalidated SAM-based ground truth and missing evaluation details keep it from being conclusive. read the letter →

arxiv 2506.11126 v1 pith:V2KDYSV4 submitted 2025-06-10 cs.CV

Artem Solomko , Oleg Kartashev , Andrey Golov , Mikhail Deulin , Vadim Valynkin , Vasily Kharin This is my paper
classification cs.CV
keywords segmentationinstanceironorepelletscomputervisionmeasuringstar-convexpolygonsStarDistWassersteindistance
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 reports that a segmentation method developed for microscopy cell nuclei, StarDist, which represents each object as a star-convex polygon, transfers successfully to industrial quality control of iron ore pellets. On a deferred test set it obtains an accuracy of 0.9916 and IoU of 0.9617, with precision and recall near 0.98, outperforming Mask R-CNN and U-Net++ on the same data. It further claims that fitting a bounding circle to the segmented contours yields physical pellet diameters whose mean absolute error against two months of laboratory measurements is about 4.9 ± 0.1 mm. If these numbers hold, one 2D-camera pipeline can sort pellets into quality classes and measure the size distribution that defines product quality.

What carries the argument

The central object is the star-convex polygon: each pellet is represented by its center plus 32 radial distances to the boundary, so convex, touching objects can be separated without bounding boxes. It is carried by a StarDist-style network with three outputs (distance, star-ray, and type) whose weighted masked losses are summed during training, followed by Non-Maximum Suppression and morphological smoothing of the predicted masks. Two supporting mechanisms matter: a Wasserstein-distance-based pixel-stratified dataset split, which keeps rare classes balanced by pixel area rather than instance count, and a semi-automatic annotation loop in which a general segmentation model's outlines are expanded by label-aware morphological operations to compensate for its tendency to underestimate pellet boundaries.

What would settle it

Re-annotate a random subset of the deferred test images by hand, without any morphological expansion, and recompute IoU, precision, and recall; if the numbers drop substantially, the reported metrics partly measure label-expansion artifacts rather than true pellet boundaries. Separately, place a calibrated physical reference directly under the camera on the same pellets that were imaged and compare per-pellet diameters, which would test the mean absolute error of about 4.9 mm without the current location mismatch.

Watch

Extended reading notes

Core claim

The central discovery is that star-convex polygon regression, originally designed for densely packed cell nuclei, handles touching and partially overlapping iron ore pellets better than general-purpose instance segmenters. The pipeline uses an EfficientNetV2-b0 backbone with three output heads: a distance map for object presence, a 32-ray star-convex map for contours, and a type map for the four classes 'nice', 'joint', 'ugly', and 'big'; the weighted masked losses from each head are summed, and Non-Maximum Suppression over predicted star polygons separates clumped objects. The authors find that a train/test split stratified by pixel-class distribution using the Wasserstein distance raises IoU from 0.9343 to 0.9617 compared with an instance-based split, and that morphological expansion of semi-automatic annotations corrects a systematic boundary underestimation. The resulting masks, converted to physical dimensions by a bounding circle fitted to contours of at least eight points, reproduce laboratory size measurements with a mean absolute error of about 4.9 ± 0.1 mm.

Load-bearing premise

The reported numbers rest on assuming that the semi-automatic annotations, after a fixed morphological expansion, are correct and unbiased, and that pellets hand-collected from a separate location fairly represent what the camera sees.

Editorial extensions

If this is right

  • A single 2D camera plus the star-convex pipeline can replace manual laboratory size-distribution analysis for iron ore pellets, with mean absolute diameter error near 5 mm.
  • On this industrial dataset, the star-convex approach beats Mask R-CNN by roughly 2.8 IoU points and U-Net++ by roughly 1.2 points, while avoiding the fused-mask problem observed with U-Net++.
  • Using a pixel-stratified Wasserstein-based split instead of an instance-based split raises final IoU from 0.9343 to 0.9617, showing that dataset splitting strategy materially changes measured performance on imbalanced industrial data.
  • The model is selected on the 'ugly' class because that class signals production violations, so the reported metrics reflect a system tuned to detect process dirt rather than only to maximize average accuracy.
  • Precise segmentation masks are presented as a reusable feature source for downstream process-analysis algorithms, such as estimating process humidity or flagging technology violations.

Reading between the lines

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

  • The paper does not decompose the ~5 mm laboratory error into camera calibration, contour-fitting, and sampling-location components; a natural next step is to place a calibrated reference sensor directly under the camera on the same pellets that are imaged and compare per-pellet diameters.
  • Because the star-convex mechanism is shape-generic, a plausible extension is to test the same pipeline on other round granular materials, such as tablets, seeds, or gravel; the paper gives no evidence beyond pellets but the representation does not depend on pellet chemistry.
  • The reported IoU gain could partly reflect the morphological expansion applied to SAM-generated ground truth, so sweeping the expansion radius on a manually re-annotated subset would separate true algorithmic performance from label-engineering artifacts.
  • Selecting the best model on the 'ugly' class means boundary precision on common classes may have been traded for rare-class sensitivity; a cost-weighted evaluation would clarify whether the 0.9617 IoU persists when false negatives on defects are penalized more heavily.
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

4 major / 6 minor

Summary. The paper proposes an image-based pipeline for instance segmentation, classification, and physical dimension measurement of iron ore pellets on a conveyor line. The method adapts StarDist with an EfficientNetV2 backbone, trained on ground-truth masks produced semi-automatically with SAM and then expanded by morphological operations, with a Wasserstein-distance-based stratified train/test split. The authors report a test-set IoU of 0.961676, precision/recall of 0.98039/0.98021 for their StarDist-based model, claim superiority over Mask R-CNN and U-Net++, and report a laboratory validation with mean absolute error of about 4.9 ± 0.1 for pellet dimensions.

Significance. If the reported results are robust, the paper demonstrates a useful cross-domain transfer: applying a cell-nuclei segmentation method (StarDist) to densely packed industrial objects with smoothed boundaries, while also contributing a Wasserstein-distance-based stratified splitting procedure and a morphological correction strategy for SAM-based annotations. The inclusion of three baseline comparisons (Mask R-CNN, U-Net++, ViT-based approaches) and an external laboratory check is a strength, as is the explicit acknowledgment that the laboratory samples came from a different physical location than the camera view. However, the significance is currently tempered by the absence of uncertainty estimates, per-class metrics, a stated IoU matching threshold, and independent verification of the ground-truth boundary correction; the central quantitative claims are plausible but not yet fully evidenced.

major comments (4)
  1. [§3.1] The ground-truth masks are generated by SAM and then expanded by morphological operations, but the paper provides no independent verification that the corrected boundaries coincide with true pellet edges. Because the same masks are used for training and for computing all reported metrics, any systematic boundary offset is learned into the target and can inflate the reported IoU and precision. I recommend annotating a subset of images manually or by an independent expert, reporting boundary agreement (e.g., boundary IoU or Hausdorff distance) per class, and reevaluating the model on the corrected labels.
  2. [§3.3 and Table 2] The metric definition in §3.3 states that a predicted object is a true positive only if its IoU with a ground-truth object exceeds a threshold τ, but the value of τ is never reported. In addition, the results in Table 2 are single point estimates with no variance, no per-class breakdown, and no statement of the test-set size (number of images and objects). The 'accuracy' value of 0.991564 in §4 is also undefined (pixel accuracy versus instance accuracy). Please specify τ, define all metrics, report means and standard deviations over multiple runs, and give per-class IoU/precision/recall along with the train/test split sizes.
  3. [§5] The laboratory validation reports a Mean Absolute Error of 4.9 ± 0.1 without units, and the authors state that laboratory samples were collected from a specially designated location rather than directly under the camera. The model and the laboratory measurements therefore concern different populations, no statistical test is performed, and the MAE does not validate per-object boundary accuracy or the pixel-to-millimeter calibration. Please state the units, describe the calibration procedure, and either perform a matched comparison on the same physical pellets or explicitly restrict the claim to distribution-level agreement rather than per-pellet measurement accuracy.
  4. [Table 2 and §4] The comparison with Mask R-CNN and U-Net++ does not state whether the baselines were trained on the same folds, the same image resolution, the same matching threshold, or the same hyperparameter selection procedure. Since the best model was selected based on the 'ugly' class, per-class metrics are necessary to support the claimed superiority of the StarDist-based approach by roughly 2 to 3 IoU points. Please document the full evaluation protocol for all compared methods and provide per-class results for each method.
minor comments (6)
  1. [Table 2] The table heading contains a typo ('Metrcis'), and the caption of Fig. 11 contains 'comprasion'; please correct these and other spelling errors throughout the manuscript.
  2. [Equation (1)] The expression for the Wasserstein distance is garbled and incomplete; please give a standard, well-defined definition and explain concretely how it was used to perform the pixel-stratified dataset split.
  3. [§2] The phrase 'pyramid patch weight smoothing' is not defined anywhere in the paper; either describe the operation precisely or remove the term.
  4. [§6] The passage beginning 'The statement refers to a process of normalizing the luminance channel...' reads like an inserted boilerplate definition and should be rewritten as an integrated part of the method description.
  5. [Figures 3 and 4] Figures 3 and 4 are referenced but not fully explained; please add legends, axis labels, and a clear description of what each output branch (distance, type, stardist) displays.
  6. [§4] The reported metrics are given with excessive significant digits (e.g., 0.961676); please report a reasonable number of digits consistent with the actual precision of the measurements.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: supervised training and deferred-set evaluation are externally grounded; reported metrics and lab MAE do not reduce by construction to model inputs.

full rationale

The paper is an empirical supervised-learning study, not a derivation from first principles. The central results (Section 4: IoU 0.961676, precision 0.98039, recall 0.98021 on a deferred dataset; Section 5: lab MAE ~4.9 ± 0.1) are obtained by training a StarDist-based model on annotated images and evaluating it on held-out data. The ground-truth masks were produced semi-automatically with SAM and then expanded by morphological operations (Section 3.1); this is a data-quality and label-fidelity risk, not a circularity, because the model's predictions are not constructed from the test labels and the reported numbers are not equations that equal their own inputs. The comparison against Mask R-CNN and U-Net++ uses the same benchmark protocol, so the claimed advantage is an empirical outcome rather than a definitional one. The paper explicitly acknowledges a limitation in the physical validation: 'Statistical hypotheses were not tested because the laboratory samples were collected from a specially designated location rather than directly under the camera' (Section 5), which reduces confidence in the MAE comparison but does not make it circular, since laboratory measurements are external to the model. No load-bearing self-citations were found; references to StarDist [12,18] are external prior work, not arguments whose conclusion is the present claim. Loss weights, ray count, and post-processing choices are hyperparameters rather than quantities fitted to the reported evaluation metrics. Therefore no step in the claimed chain reduces by definition, by fit, or by a self-citation loop.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities. Its central claim rests on the quality of semi-automatic annotations, the stability of the four-class taxonomy, and the geometric assumption that a bounding circle measures pellet size. The free parameters listed are hyperparameters and reporting gaps rather than physics-style fitted constants.

free parameters (5)
  • Loss weights (distance, type, stardist) = 1.0, 1.0, 0.5
    Chosen by hand in Equation 2 to emphasize class and distance branches; not fitted to the evaluation data, but they affect the final model and metrics.
  • Number of StarDist rays = 32
    Set in Section 3.2 for the star-convex polygon representation; a higher ray count would change contour fidelity and measured dimensions.
  • Training hyperparameters = learning rate 3e-4, weight decay 1e-5, 1000 epochs, MultiStepLR at 500 and 800
    Reported in Section 3.2; these are standard choices, but they affect the final performance and are not swept or justified.
  • IoU matching threshold tau = not reported
    Section 3.3 defines true positives by IoU greater than a threshold tau in [0,1], but the actual threshold used in Table 2 is never stated, making the metrics non-reproducible.
  • MAE value and its uncertainty = 4.9 ± 0.1 (units unspecified)
    The dimension-validation result in Section 5 has no units and no description of how the ±0.1 error was computed, so its meaning is ambiguous.
assumptions (4)
  • domain assumption SAM-based semi-automatic annotations, after morphological expansion, accurately represent true pellet boundaries and classes.
    Section 3.1 relies on SAM for about 80% of the annotations and applies expanding labels to fix boundary underestimation; all training and evaluation assume this corrected ground truth is correct.
  • domain assumption The four-class taxonomy (nice, joint, ugly, big) is well-defined and can be applied consistently by human annotators.
    Table 1 gives subjective visual descriptions (round, oval, rod-shaped, adhering elements) without quantitative criteria; the classifier and the evaluation metrics depend on the consistency of these labels.
  • domain assumption A bounding circle fitted to the segmented contour is an adequate proxy for physical pellet diameter.
    Section 3.2 states 'the diameter of such a circle was taken as the final pellet size,' and Section 4 states measurements are made only on nice pellets. No calibration procedure or uncertainty analysis is given for this geometric simplification.
  • domain assumption Luminance normalization of input images aligns them with the training distribution and improves generalization.
    Section 6 introduces CIELAB luminance normalization and hard brightness augmentations, assuming the training luminance distribution is representative of deployment conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Image-Based Method For Measuring And Classification Of Iron Ore Pellets Using Star-Convex Polygons." pith.science (2026). https://pith.science/paper/V2KDYSV4

@misc{pith2026250611126,
  author       = {Pith},
  title        = {Pith review of: Image-Based Method For Measuring And Classification Of Iron Ore Pellets Using Star-Convex Polygons},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V2KDYSV4}},
  note         = {Machine review of arXiv:2506.11126}
}
read the original abstract

We would like to present a comprehensive study on the classification of iron ore pellets, aimed at identifying quality violations in the final product, alongside the development of an innovative imagebased measurement method utilizing the StarDist algorithm, which is primarily employed in the medical field. This initiative is motivated by the necessity to accurately identify and analyze objects within densely packed and unstable environments. The process involves segmenting these objects, determining their contours, classifying them, and measuring their physical dimensions. This is crucial because the size distribution and classification of pellets such as distinguishing between nice (quality) and joint (caused by the presence of moisture or indicating a process of production failure) types are among the most significant characteristics that define the quality of the final product. Traditional algorithms, including image classification techniques using Vision Transformer (ViT), instance segmentation methods like Mask R-CNN, and various anomaly segmentation algorithms, have not yielded satisfactory results in this context. Consequently, we explored methodologies from related fields to enhance our approach. The outcome of our research is a novel method designed to detect objects with smoothed boundaries. This advancement significantly improves the accuracy of physical dimension measurements and facilitates a more precise analysis of size distribution among the iron ore pellets. By leveraging the strengths of the StarDist algorithm, we aim to provide a robust solution that addresses the challenges posed by the complex nature of pellet classification and measurement.

Figures

Figures reproduced from arXiv: 2506.11126 by the authors.

Figure 1
Figure 1. The 2D-camera’s region of interest is represented in the image – densely spaced objects make it difficult to separate them into classes. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Image of a cells with nuclei (left) and pellets (right). 2 Method We implemented the classic StarDist calculation method, which is a well-established approach in the field of object detection and segmentation. This method is distinctive because it allows us to predict the shapes of objects of interest directly, without the need for traditional bounding boxes. Instead, we utilize a contour-based representation known … view at source ↗
Figure 3
Figure 3. Model diagram. The algorithm assumes the return of several types of outputs: distance, type, and stardist. Each branch is trained with a separate loss function, and the final loss, which is used for the optimizer step, is the sum of the losses of all branches of the algorithm. The distance maps are used as the object probability to determine whether the model detects objects as pellets or not. The stardist maps are … view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Different masks-output from algorithm. For the “stardist” output, we set the masked MAE. For the “distance” outputs, we use the masked BCELoss combined with MSE, and for the “type” output — the masked multiclass categorical CELoss combined with DiceLoss. The following …
Figure 5
Figure 5. Figure 5: FastFlow. The interpretation of such masks is difficult and does not allow solving the problem with one model. Therefore, we explored additional approaches. A metric learning approach for classification was carried out by dividing frames into two classes: ugly and nice…
Figure 6
Figure 6. Figure 6: Instance segmentation masks by Segment Anything method w/o tuning. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Stardist detections vs maskRCNN. The colors reflect the different classes. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Unet++ preds. The colors reflect the different classes. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 8
Figure 8. Figure 8: Unet++ preds. The colors reflect the different classes (continue). [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Instance segmentation masks on images by StarDists. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Contours of pellets. Several approaches were used for contours: circle, convex hull, ellipse. The approach using circles proved to be the best [1]. Measurements are made only on “nice”. pellets, so this form conveyed the physical dimensions of the object more accurate…
Figure 11
Figure 11. Figure 11: Visual comprasion of different approaches: b0 w/o wassertein, b1 with wasserstein, b0 with wasserstein. The colors reflect the different classes. [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Comparison with laboratory experiments. Black is our model, blue-laboratory. [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Corrupted detections. The statement refers to a process of normalizing the luminance channel of an image to align it with a specific standard within a dataset, specifically using the CIELAB color space. In this context, the luminance channel represents the brightness …
Figure 14
Figure 14. Figure 14: Fixed detections. This approach enhances the model’s predictions by increasing their consistency and reliability ( [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Dependence on luminance. 7 Conclusion The initial objective of our project was to identify and categorize the various types of pellets. However, as we progressed, the capabilities of our algorithm evolved, allowing us to incorporate additional functionalities, such as…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

  1. [1]

    Instancecut: from edges to instances with multicut

    Kirillov A, Levinkov E, Andres B, Savchynskyy B, and Rother C. Instancecut: from edges to instances with multicut. 2016

    work page 2016

  2. [2]

    Segment anything

    Kirillov Alexander, Mintun Eric, Ravi Nikhila, Mao Hanzi, Rolland Chloe, Gustafson Laura, Xiao Tete, Whitehead Spencer, Berg Alexander, Lo Wan-Yen, Dollar Piotr, and Girshick Ross. Segment anything. 2023

    work page 2023

  3. [3]

    Panoptic segmentation

    Kirillov Alexander, He Kaiming, Girshick Ross, Rother Carsten, and Dollar Piotr. Panoptic segmentation. 2019

    work page 2019

  4. [4]

    Complete scanning application using opencv

    Gangal Ayushe, Kumar Peeyush, and Kumari Sunita. Complete scanning application using opencv. 2021

    work page 2021

  5. [5]

    Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows

    Yu Jiawei, Zheng Ye, Wang Xiang, Li Wei, Wu Yushuang, Zhao Rui, and Wu Liwei. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. 2021

    work page 2021

  6. [6]

    Maskrcnn"

    He Kaiming, Gkioxari Georgia, Dollar Piotr, and Girshick Ross. Maskrcnn". 2018

    work page 2018

  7. [7]

    Topological data analysis

    Wasserman L. Topological data analysis. 2018

    work page 2018

  8. [8]

    Image-based method for measuring pellet size distribution in the stable area of disc pelletizer

    Xiaoyan Liu, Chuangang Mao, Sun Wei, and Xin W. Image-based method for measuring pellet size distribution in the stable area of disc pelletizer. 2018

    work page 2018

Show all 19 references
  1. [9]

    Spatial transformer networks

    Jaderberg M, Simonyan K, Zisserman A, and Kavukcuoglu K. Spatial transformer networks. 2015

    2015

  2. [10]

    Statistical aspects of wasserstein distances

    Victor M, Panaretos Yoav, and Zemel. Statistical aspects of wasserstein distances. 2018

    2018

  3. [11]

    Nuclei instance segmentation and classification in histopathology images with stardist

    Weigert Martin and Schmidt Uwe. Nuclei instance segmentation and classification in histopathology images with stardist. 2022

    2022

  4. [12]

    Star-convex polyhedra for 3d object detection and segmentation in microscopy

    Weigert Martin, Schmidt Uwe, Haase Robert, Sugawara Ko, and Myers Gene. Star-convex polyhedra for 3d object detection and segmentation in microscopy. 03 2020

    2020

  5. [13]

    Efficientnetv2: Smaller models and faster training

    Tan Mingxing, Quoc V , and Le. Efficientnetv2: Smaller models and faster training. 2021

    2021

  6. [14]

    Morphological image analysis: Principles and applications

    Soille P. Morphological image analysis: Principles and applications. 1999

    1999

  7. [15]

    Lefuvre S. Int. conf. on computer analysis of images and patterns. 2007

    2007

  8. [16]

    Conic: Colon nuclei identification and counting challenge 2022

    Graham Simon, Jahanifar Mostafa, Vu Quoc, Hadjigeorghiou Giorgos, Leech Thomas, and Snead David. Conic: Colon nuclei identification and counting challenge 2022. 2021

    2022

  9. [17]

    Focal loss for dense object detection

    Lin Tsung-Yi, Goyal Priya, Girshick Ross, He Kaiming, and Dollar Piotr. Focal loss for dense object detection. 2017

    2017

  10. [18]

    Cell detection with star-convex polygons

    Schmidt Uwe, Weigert Martin, Broaddus Coleman, and Myers Gene. Cell detection with star-convex polygons. 2018

    2018

  11. [19]

    Unet++: A nested u-net architecture for medical image segmentation

    Nima Tajbakhsh Jianming Liang Zongwei Zhou, Md Mahfuzur Rahman Siddiquee. Unet++: A nested u-net architecture for medical image segmentation. 2018. 15

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.