Pith. sign in

REVIEW 4 major objections 5 minor 23 references

Creation and Evaluation of a Food Product Image Dataset for Product Property Extraction

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper builds a 1,034-image dataset of 250 food products with GS1-aligned labels and baselines, offered as a first-step resource for retail product recognition and property extraction.

desk verdict A modest but genuine GS1-aligned food-product image dataset with honest baselines; the main gap is unquantified annotation reliability, which the authors themselves concede. read the letter →

arxiv 2411.10591 v1 pith:4GWR7X7Y submitted 2024-11-15 cs.CV cs.LG

Christoph Brosch , Alexander Bouwens , Sebastian Bast , Swen Haab , Rolf Krieger This is my paper · ORCID
classification cs.CVcs.LG
keywords foodproductimagedatasetpropertyextractionobjectdetectionclassificationGS1standardsretailcomputervisiongroundtruthannotationbaselinemodels
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

This paper reports the construction of a food product image dataset meant to support machine-learning systems in retail: 1,034 studio images of 250 single products, annotated with five image-level labels, 30 object-detection label classes (barcode, nutrition table, brand name, and similar), and ground-truth property values such as fill weight, alcohol content, and seal types. The authors argue that existing single-product image datasets are not enough for their purpose, because they lack legible English or German text, professional background and lighting, and detailed region annotations. Their dataset follows GS1 standards for product classification, product image capture, packaging codes, and label naming, which they say makes it easy to extend. They include unmodified baseline models, ResNet50 for image classification and YOLOv5 for object detection, so future work has concrete numbers to improve on. The central claim is that the resource supports training and evaluation of product recognition, classification, and property extraction models.

What carries the argument

The carrying mechanism is the dataset schema itself, organized around GS1 standards: the Global Product Classification hierarchy sets which products to include and how to classify them, the GS1 product image specification dictates the capture setup, the GS1 packaging codes supply packaging-type labels, and the GS1 web vocabulary supplies label naming conventions. Around this shell sits a 30-class object detection taxonomy grounded in EU Regulation 1169/2011, stored in Pascal VOC format with optional difficult and truncated flags, plus five image-level classification attributes and structured product property values. That schema is what lets the dataset work as an end-to-end testbed, connecting pixel-level bounding boxes to machine-readable product facts. The baseline models are deliberately generic ResNet50 and YOLOv5 networks, whose role is to provide a comparison floor rather than to break new algorithmic ground.

What would settle it

Re-annotate a random subset of the 250 products with independent annotators following the same label list and compare their bounding boxes and class labels, and also check the extracted property values against what is legible on the packages; if agreement between annotators is low or many property values are wrong, the claim that the dataset provides reliable ground truth would fail.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a resource plus a reproducible process, not a new algorithm. The dataset pairs each photograph with an XML file of Pascal VOC bounding boxes for 30 packaging elements, a CSV of image attributes (packaging type, material, facing, fill type), and a CSV of product properties including nutrition data, GTIN, brand, and GPC brick. The labels were chosen by combining the GS1 web vocabulary with what European Regulation 1169/2011 requires on food packaging, and images were captured according to the GS1 product image specification. The authors claim this combination fills a gap left by shelf-scene and rotated-plate datasets by providing clean lighting, neutral background, and readable text, and they present baseline results that let later work measure improvement.

Load-bearing premise

Everything the paper offers as a benchmark resource depends on the hand-made labels and extracted property values being accurate enough to be treated as ground truth, even though the authors themselves flag that human error and varying interpretations may have introduced inconsistencies.

Editorial extensions

If this is right

  • Researchers can train and evaluate object detectors on 30 packaging elements in clean, single-product images, with per-label difficult and truncated flags indicating occlusion or poor legibility.
  • The paired CSV property values allow end-to-end evaluation of a pipeline that detects a region, reads its text, and fills a structured product record.
  • Because selection and labels follow GS1, other groups can extend the dataset with new products while keeping the same schema and naming conventions.
  • The reported baseline precision, recall, and mAP numbers give a concrete reference point that later models can be compared against.
  • The paper's workload estimates (about 5 minutes of photography and 20 minutes of annotation per product) let others plan the cost of enlarging the set.

Reading between the lines

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

  • The difficult and truncated flags could be used to quantify how much occlusion and reflection actually hurt detection, since the paper reports aggregate metrics without separating flagged from unflagged boxes.
  • The pairing of bounding boxes with structured property values suggests a natural next task the paper does not evaluate: matching detected text regions to specific product fields, which would test the full property-extraction loop.
  • The paper's early result that synthetic images did not transfer to the real detection model implies that simple data augmentation is unlikely to solve the rare-label problem; a more targeted test would be to measure per-label learning curves against the number of real images.
  • The acknowledged possibility of human label errors means an inter-annotator agreement study on a random subset would be the cheapest check of whether the dataset can truly serve as ground truth.
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 / 5 minor

Summary. The paper describes the creation and evaluation of a food-product image dataset containing 1,034 studio photographs of 250 single food products. Each image is accompanied by Pascal VOC bounding boxes for 30 packaging-property classes (e.g., productName, barcode, nutritionTable), five image-level classification labels (packaging type, material, facing, fill type, and in/out-of-packaging status), and a product-information CSV with 30 attributes including 15 nutritional values. The selection and annotation process follows GS1 standards (GPC, product-image specification, packaging codes, and web vocabulary). The authors report label statistics and baseline results: a ResNet50 fine-tuned head for three classification tasks and a YOLOv5 detector for the 30 object classes. They argue that the dataset is an important first step and a useful resource for product recognition and property extraction, with a documented process for extension.

Significance. If the annotation quality can be confirmed, this is a useful resource for the retail-computer-vision community: it is GS1-aligned, publicly available, and combines localization labels with property-level ground-truth text, which existing datasets such as RPC and Products-6k do not offer. The authors ship notebooks, scripts, and baseline configurations, and the baseline experiments are standard and reproducible in principle. The main risk is the unvalidated accuracy of the manual annotations: Section 4 acknowledges possible inconsistencies, no inter-annotator agreement is reported, and Table 3's metrics are computed on the same labels. This is not a circularity problem, because evaluating a dataset with baselines trained on it is standard practice; it is an annotation-quality problem that must be addressed before the numerical claims can be used as benchmarks.

major comments (4)
  1. [4, Image Annotation, and Section 4.4] The paper's central claim that the dataset is a valuable benchmark rests on the accuracy of the hand-created object labels and the extracted property values. Section 4 explicitly concedes that "inconsistencies may arise in our labelled data, due to varying interpretations of the specifications and human errors," but the paper provides no inter-annotator agreement, no spot-check protocol, and no external validation of the Product Information CSV (30 attributes, 15 nutrition values) described in Section 4.4. Since Table 3 reports object detection baselines computed on exactly these annotations, the mAP numbers inherit any label noise. Please report the number of annotators, the annotation instructions, agreement statistics on a re-annotated subset (per label or at least for the bounding-box classes), and the source and verification procedure for the property-value CSV.
  2. [6, Table 3] The object detection baselines are reported as single numbers from one training run, without confidence intervals, standard deviations, or a description of how hyperparameters and random seeds were chosen ("default parameters" only). Given the small dataset (1,034 images, 200 training products / 50 validation products), training variability could be non-negligible, and the paper does not state that the split is versioned or downloadable. Please run the baseline multiple times with different seeds, report mean and standard deviation for mAP@.5 and mAP@.5:.95, and make the exact train/validation split and configuration files part of the repository.
  3. [5 and Table 3] Several detection classes have extremely few instances: countryOfOrigin (13), allergenStatement (8), percentageOfAlcoholByVolume (1), hazards (1), priceSpecification (1), isFrozen (7), and hasReturnablePackageDeposit (9). For these classes the per-class mAP in Table 3 is either 0 or based on a handful of boxes (e.g., percentageOfAlcoholByVolume has recall 0, and priceSpecification has mAP@.5 0.497 from a single label), so the numbers are not meaningful as performance estimates. The aggregate mAP is correspondingly dominated by the eight or so common labels. Please report per-class results only for classes with a minimum number of instances, or provide confidence intervals, and state the threshold; otherwise the reader cannot infer which of the 30 labels are actually usable.
  4. [6, Table 2] The image-classification baseline is under-specified: Table 2 reports three tasks (material, packagingType, facing) but not the remaining two of the five advertised class labels, and no train/validation split, class lists, training epochs, or learning rate are given. The text says only that "the head of each model has been trained" on a pre-trained ResNet50. Without these details the classification baselines cannot be reproduced, and the reader cannot tell why two labels were omitted. Please provide the missing protocol or explicitly scope the baseline claim.
minor comments (5)
  1. [3.1, Table 1] The statement that each family should be represented by at least 10 products cannot be verified, because Table 1 reports only family/class/brick coverage counts, not per-family product counts; please include a supplementary table with product counts per family (and per brick where possible).
  2. [4.3] The five image-level classification labels are not explicitly enumerated as a set; listing the five labels and their value inventories would make the annotation schema easier to reuse.
  3. [5, Figure 4] Figure 4a is difficult to read because the 30 label names are densely spaced; a sorted table or a horizontal bar chart with counts would be clearer.
  4. [References] The reference to Regulation (EU) No 1169/2011 is incomplete; please add the full title and publication details (OJ series and date), and check the access dates for the GS1 web references.
  5. [Code Availability] The repository link should state a license, a version or DOI, and the exact split used for the Table 2 and Table 3 experiments; this is needed for the dataset to be citable and the baselines reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's dataset contribution and baselines are self-contained; no prediction reduces to its inputs by construction.

full rationale

The paper makes no theoretical derivation and fits no parameters from which a prediction is then claimed; its central claim is that a newly created, annotated dataset (1,034 studio images of 250 food products with 5 class labels, 30 object detection labels, and extracted property values) constitutes a useful resource, and that the reported ResNet50 and YOLOv5 results are baselines. Evaluating baselines on the same dataset is standard practice for dataset papers and does not make the dataset's value circular: the contribution is the resource itself, and the baseline numbers are explicitly framed as starting points, not as independent validation of an external hypothesis. The annotation process follows external GS1 standards and EU regulation references, and the object-detection label scheme is derived from observable packaging properties rather than from the target result. The only internal loop the reader could identify is that dataset usefulness is demonstrated by training models on that same dataset, which is a conventional use of a benchmark and not a derivation that reduces to its own inputs. The authors' candid Section 4 note that 'inconsistencies may arise in our labelled data, due to varying interpretations of the specifications and human errors' is a limitation concerning annotation reliability and ground-truth quality, not a circularity; it does not cause any claimed result to be equivalent to its inputs by construction. No self-citations are load-bearing, no fitted quantity is renamed as a prediction, and no known result is repackaged under new coordinates. Accordingly, the circularity score is 0.

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

No free parameters are fitted and no new theoretical entities are introduced. The paper's claims instead rest on domain assumptions: that GS1 standards provide a valid annotation and selection basis, that the 30-label set covers relevant packaging properties, that the captured studio images show all relevant product sides, and that human annotations are consistent enough to be used as ground truth. None of these assumptions is independently validated in the paper.

assumptions (4)
  • domain assumption GS1 standards (GPC, product image specification, packaging codes, web vocabulary) provide a valid and sufficient basis for selecting products and defining annotations.
    The dataset design is built on GS1 standards cited in Sections 3.1 and 4.1; if those standards are incomplete or inappropriate for the target property extraction task, the annotation schema and product coverage would be weakened.
  • domain assumption The set of 30 object detection labels, derived from common German packaging elements and EU Regulation 1169/2011, covers the relevant product properties for extraction.
    Section 4.2 derives this label set without evidence of completeness; rare but important properties could be missing, limiting the dataset's usefulness.
  • domain assumption The manually created annotations and ground-truth property values are consistent enough to be used as training and evaluation labels.
    Section 4 begins with an admission that inconsistencies may arise from human error; no inter-annotator agreement or independent validation is reported, so this reliability is assumed.
  • domain assumption Studio images with white background and two to six straight-on views per product are sufficient to support product property extraction.
    Section 3.3 describes the capture protocol; its adequacy for OCR and end-to-end property extraction is asserted rather than tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Creation and Evaluation of a Food Product Image Dataset for Product Property Extraction." pith.science (2026). https://pith.science/paper/4GWR7X7Y

@misc{pith2026241110591,
  author       = {Pith},
  title        = {Pith review of: Creation and Evaluation of a Food Product Image Dataset for Product Property Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4GWR7X7Y}},
  note         = {Machine review of arXiv:2411.10591}
}
read the original abstract

The enormous progress in the field of artificial intelligence (AI) enables retail companies to automate their processes and thus to save costs. Thereby, many AI-based automation approaches are based on machine learning and computer vision. The realization of such approaches requires high-quality training data. In this paper, we describe the creation process of an annotated dataset that contains 1,034 images of single food products, taken under studio conditions, annotated with 5 class labels and 30 object detection labels, which can be used for product recognition and classification tasks. We based all images and labels on standards presented by GS1, a global non-profit organisation. The objective of our work is to support the development of machine learning models in the retail domain and to provide a reference process for creating the necessary training data.

Figures

Figures reproduced from arXiv: 2411.10591 by the authors.

Figure 1
Figure 1. Example images from our dataset with visualized [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Example data for a single product image. Shows [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Example of images with difficult (left) and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Statistical analysis of our dataset. N = 1034. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 19 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    Chen, F., Zhang, H., Li, Z., Dou, J., Mo, S., Chen, H., Zhang, Y., Ahmed, U., Zhu, C., and Savvides, M. (2022). Unitail: Detecting , Reading , and Matching in Retail Scene . arXiv:2204.00298 [cs]

    work page Pith review arXiv 2022

  3. [3]

    F., Dideriksen, T., Arora, H., Guillaumin, M., and Malik, J

    Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Yago Vicente, T. F., Dideriksen, T., Arora, H., Guillaumin, M., and Malik, J. (2022). Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR

    work page 2022

  4. [4]

    LabelImg

    darrenl (2022). LabelImg . Available at https://github.com/tzutalin/labelImg, last accessed 14.03.2023

    work page 2022

  5. [5]

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. (2009). Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition , pages 248--255

    2009

  6. [6]

    REGULATION ( EU ) No 1169/2011 OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL of 25 October 2011

    European Parliament and the Council (2011). REGULATION ( EU ) No 1169/2011 OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL of 25 October 2011

    work page 2011

  7. [7]

    Everingham, M., Van Gool, L., Williams, C. K. I., Winn, J., and Zisserman, A. (2010). The Pascal Visual Object Classes ( VOC ) Challenge . International Journal of Computer Vision , 88(2):303--338

    work page 2010

  8. [8]

    Follmann, P., Böttger, T., Härtinger, P., König, R., and Ulrich, M. (2018). MVTec D2S : Densely Segmented Supermarket Dataset

    work page 2018

Show all 23 references
  1. [9]

    and Floerkemeier, C

    George, M. and Floerkemeier, C. (2014). Recognizing Products : A Per -exemplar Multi -label Image Classification Approach . In Fleet, D., Pajdla, T., Schiele, B., and Tuytelaars, T., editors, Computer Vision – ECCV 2014 , pages 440--455, Cham. Springer International Publishing

    2014

  2. [10]

    Georgiadis, K., Kordopatis-Zilos, G., Kalaganis, F., Migkotzidis, P., Chatzilari, E., Panakidou, V., Pantouvakis, K., Tortopidis, S., Papadopoulos, S., Nikolopoulos, S., and Kompatsiaris, I. (2021). Products- 6K : A Large - Scale Groceries Product Recognition Dataset . In The ...

    2021

  3. [11]

    Goldman, E., Herzig, R., Eisenschtat, A., Goldberger, J., and Hassner, T. (2019). Precise detection in densely packed scenes. In Proc. Conf. Comput. Vision Pattern Recognition (CVPR)

    2019

  4. [12]

    Global Product Classification ( GPC ) GS1

    GS1 (2020). Global Product Classification ( GPC ) GS1 . Available at https://www.gs1.org/standards/gpc, last accessed 17.03.2023

    2020

  5. [13]

    GS1 Product Image Specification Standard GS1

    GS1 (2022). GS1 Product Image Specification Standard GS1 . Available at https://www.gs1.org/standards/gs1-product-image-specification-standard/current-standard, last accessed 17.03.2023

    2022

  6. [14]

    GS1 Web Vocabulary Food Beverage Tobacco Product

    GS1 (2023). GS1 Web Vocabulary Food Beverage Tobacco Product . Available at https://www.gs1.org/voc/FoodBeverageTobaccoProduct, last accessed 11.04.2022

    2023

  7. [15]

    Codes for types of packaging - GS1 Netherlands

    GS1 Netherlands (2022). Codes for types of packaging - GS1 Netherlands . Available at https://gs1.nl/en/knowledge-base/gs1-datapools-overview/codes-for-types-of-packaging/, last accessed 17.03.2023

    2022

  8. [16]

    He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep Residual Learning for Image Recognition . In 2016 IEEE Conference on Computer Vision and Pattern Recognition ( CVPR ) , pages 770--778, Las Vegas, NV, USA. IEEE

    2016

  9. [17]

    heartexlabs/label-studio

    Heartex (2023). heartexlabs/label-studio. Available at https://github.com/heartexlabs/label-studio, last accessed 12.04.2022

    2023

  10. [18]

    Jocher, G., Nishimura, K., Mineeva, T., and Vilariño, R. (2020). yolov5. Code repository https://github. com/ultralytics/yolov5

    2020

  11. [19]

    Liu, L., Ouyang, W., Wang, X., Fieguth, P., Chen, J., Liu, X., and Pietikäinen, M. (2020). Deep Learning for Generic Object Detection : A Survey . International Journal of Computer Vision , 128(2):261--318

    2020

  12. [20]

    Merler, M., Galleguillos, C., and Belongie, S. (2007). Recognizing groceries in situ using in vitro training data. In 2007 IEEE Conference on Computer Vision and Pattern Recognition , pages 1--8

    2007

  13. [21]

    GPT -4 Technical Report

    OpenAI (2023). GPT -4 Technical Report . arXiv:2303.08774 [cs]

    2023 arXiv

  14. [22]

    Wei, X.-S., Cui, Q., Yang, L., Wang, P., and Liu, L. (2019). RPC : A Large - Scale Retail Product Checkout Dataset . Number: arXiv:1901.07249 arXiv:1901.07249 [cs]

    2019 arXiv

  15. [23]

    Wei, Y., Tran, S., Xu, S., Kang, B., and Springer, M. (2020). Deep Learning for Retail Product Recognition : Challenges and Techniques . Computational Intelligence and Neuroscience , 2020:1--23

    2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.