Pith. sign in

REVIEW 3 major objections 6 minor 18 references

Container damage detection using advanced computer vision model Yolov12 vs Yolov11 vs RF-DETR A comparative analysis

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper compares YOLOv11, YOLOv12, and RF-DETR for detecting damaged shipping containers and finds that RF-DETR identifies unusual damage more reliably despite scoring lower on the standard mAP@50 benchmark.

desk verdict A modest three-model benchmark on container damage that is honest about its quantitative results but overclaims RF-DETR superiority from three unlabeled test images. read the letter →

arxiv 2506.22517 v1 pith:4XCNFQEO submitted 2025-06-26 cs.CV

Subhadip Kumar This is my paper · ORCID
classification cs.CV
keywords containerdamagedetectionobjectYOLOv11YOLOv12RF-DETRmAPprecisionrecall
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 compares three recent object detectors—YOLOv11, YOLOv12, and RF-DETR—on the task of finding damaged shipping containers. Using 278 annotated container images, the author trains all three models and finds a mixed result: YOLOv11 and YOLOv12 reach an mAP@50 of 81.9% while RF-DETR trails at 77.7%. When the models are then tested on three unusual damage cases—a fire-damaged container, two stacked fire-damaged containers, and a bent container—RF-DETR detects more damaged sections with higher confidence than the other two. The paper argues that RF-DETR is therefore better suited for spotting not-so-common container damage, even though it loses on the standard benchmark metric.

What carries the argument

The comparison is carried by three model families and a fixed evaluation protocol. YOLOv11 and YOLOv12 are CNN-based single-shot detectors: YOLOv11 uses C3k2 blocks in the neck and head, while YOLOv12 introduces area attention and residual efficient layer aggregation networks (R-ELAN). RF-DETR replaces the anchor-based head with a transformer encoder-decoder that directly predicts boxes and classes, built on a vision-transformer backbone. The protocol trains all models on the same 234-image training set, validates on 25 images, and evaluates on 19 test images using mAP@50, precision, and recall, followed by three hand-picked out-of-distribution images meant to represent rare damage.

What would settle it

Run all three models, at identical confidence thresholds, on a pre-specified set of 50 or more rare damage container images (fire, bent, crushed) not used for training, and compute per-class recall and detection confidence; if RF-DETR does not consistently detect more damaged sections than YOLOv11/12, the central claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that RF-DETR, a transformer-based detector, outperforms YOLOv11 and YOLOv12 when the input is a not-so-common damage type. On three out-of-dataset test images—a fully fire-damaged container, two stacked fire-damaged containers, and a fully bent container—RF-DETR localized more damaged sections with confidence above 90% in the first case, 93% in the second, and three sections at 54%, 55%, and 81% in the third. YOLOv11 and YOLOv12 each missed sections or reported lower confidence on the same images. The author interprets this as RF-DETR having stronger generalization to unusual damage, despite its lower mAP@50 score on the 19-image test set.

Load-bearing premise

The claim that RF-DETR is better on not-so-common damage rests on the three hand-picked images being representative of such damage and the detections reflecting true capability rather than chance.

Editorial extensions

If this is right

  • If RF-DETR's generalization advantage is real, safety-critical container inspection should prefer transformer-based detectors over CNN-based ones for rare damage types, even when standard mAP is lower.
  • Standard mAP@50 alone would have picked YOLOv11 as the winner; the qualitative tests show that benchmark ranking can mislead deployment choices for out-of-distribution damage.
  • The mixed result implies that no single model dominates: YOLOv11 has the highest precision (91.9%), YOLOv12 the highest recall (78.4%), and RF-DETR the highest confidence on unusual cases.
  • The paper's answer to RQ2 is yes: all three deep learning models are strong enough to be practical alternatives to manual container inspection.

Reading between the lines

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

  • Three images are far too few to establish a reliable ranking for rare damage; a reader should treat the RF-DETR advantage as a promising signal, not a settled result.
  • The qualitative tests compare detection confidence and section counts, not localization accuracy against ground truth; RF-DETR's high confidence could coexist with worse bounding-box precision on a larger sample.
  • A natural follow-up is to use RF-DETR to pseudo-label rare-damage images and fine-tune YOLOv11/12 on them, potentially combining strong mAP with better coverage of unusual damage.
  • Before deployment, the models should be tested on a pre-registered set of rare damage images with repeated runs and statistical comparison, since the current protocol cannot distinguish true capability from chance.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript presents a comparative benchmark of three recent object detectors—YOLOv11, YOLOv12, and RF-DETR—for shipping container damage detection. The authors train all three models on a dataset of 278 annotated images (234/25/19 train/validation/test split) and report in Table 1 that YOLOv11 and YOLOv12 achieve 81.9% mAP@50, while RF-DETR achieves 77.7%, with precision and recall also tabulated. The paper then tests the models on three additional images (a fire-damaged container, two stacked fire-damaged containers, and a bent container) and, based on the number of predicted damaged sections and their confidence scores, concludes that RF-DETR outperforms the other two models for 'not-so-common' damage types. This conclusion is in tension with the quantitative table and is not supported by the qualitative evidence as presented.

Significance. If the results were fully supported, this comparison would be a useful practical reference for selecting a damage-detection model in container logistics, and the choice of three recent architectures (YOLOv11, YOLOv12, RF-DETR) is timely. The application domain is genuinely important, and the paper makes a concrete attempt to provide both quantitative and qualitative comparisons. However, the study is not reproducible as written: no dataset link, no training hyperparameters, no inference details, no error bars, and no ground-truth annotations for the qualitative tests are provided. The central 'RF-DETR superiority' claim rests on a three-image visual comparison that directly contradicts the reported mAP. As it stands, the work does not meet the evidentiary standard for a comparative study.

major comments (3)
  1. [§6 Conclusion and Abstract vs. Table 1] The abstract and Section 6 claim that RF-DETR 'outperformed the others overall' and exhibited 'superiority' for not-so-common damaged containers, but this is directly contradicted by Table 1, where RF-DETR has the lowest mAP@50 (77.7%) and the lowest recall (73%), while YOLOv11 and YOLOv12 both reach 81.9% mAP@50. Section 5 itself states that 'there is no clear winner,' which is inconsistent with the abstract and conclusion. The headline claim is therefore not supported by the paper's own quantitative evidence.
  2. [§5, Tests 1–3] The three-image evaluation is not a valid accuracy comparison. No ground-truth annotations are provided for the fire-damaged, stacked, or bent container images; the paper only reports the number of predicted sections and confidence values (e.g., Test 3: RF-DETR detects 3 sections at 54%, 55%, and 81%). Without ground truth, counting detections above a 50% confidence threshold cannot establish detection accuracy—the extra boxes may be false positives. Additionally, no sampling protocol or selection criterion is given for these three images, so the qualitative conclusion cannot be generalized to 'not-so-common' container damage.
  3. [§3 and §5, Table 1] The study is missing essential reproducibility and statistical information. There is no dataset identifier or link, no description of image sources or annotation protocol, no training hyperparameters (epochs, batch size, input resolution, optimizer, learning rate), no inference hardware, and no confidence intervals or repeated trials. With only 19 test images, the difference between 81.9% and 77.7% mAP@50 may be within run-to-run noise, and the paper does not report any measure of variance or significance. Without this information, neither the quantitative comparison nor the qualitative claim can be independently verified.
minor comments (6)
  1. [Throughout] There are numerous typographical and grammatical errors, for example 'overtime' in the abstract should be 'over time,' 'prolonged expose' should be 'prolonged exposure,' 'fully bend container' should be 'fully bent container,' and 'displayed showed strong performance' is redundant.
  2. [§2.3 and §2 introduction] The text describes RF-DETR as a 'single-shot object detector' based on CNN in the introduction to Section 2, but RF-DETR is a transformer-based detector built on DINOv2 and Deformable DETR; this is not a CNN single-shot model. Please clarify the architectural distinction.
  3. [§4.4] The mAP@0.5 formula is malformed: the summation index is not defined, the notation 𝛰𝜋(AP𝑢) is not standard, and the equation rendering is broken. A correct definition of average precision and the averaging procedure should be provided.
  4. [Figures 10–12] Figures 10–12 are referenced as showing the qualitative test results, but the images appear not to be included in the submitted text, so the reader cannot visually inspect the detections that support the qualitative claims.
  5. [Table 1] Table 1 has no caption and does not state whether the reported mAP, precision, and recall values are from the best epoch, the last epoch, or a selected checkpoint; this should be specified.
  6. [References] Several references are incomplete or informal, such as [2] (arXiv-only, no version), [4] (no publication venue details), and [5] (GitHub URL without a formal citation); please format the bibliography consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark with no derivation chain that could reduce to its inputs.

full rationale

This manuscript is an empirical comparison of three off-the-shelf object detectors (YOLOv11, YOLOv12, RF-DETR) on a container-damage dataset. There is no analytical derivation, no fitted parameter that is later renamed as a prediction, and no load-bearing self-citation. The quantitative results (Table 1: mAP@50, precision, recall) are reported as direct measurements on a held-out test set, and the qualitative Tests 1–3 are additional evaluations on images outside the training set. The claim that RF-DETR 'outperformed the others overall' on 'not-so-common damaged containers' rests on visual/confidence-based observations without ground-truth labels, which is a validity or evidentiary concern, not a circularity: the conclusion is not equivalent to the training objective, the metric definitions, or any input assumption. The paper even acknowledges the tension by stating that YOLOv11 and YOLOv12 had higher mAP and that 'there is no clear winner.' Since the central results are direct empirical measurements from model outputs, the derivation chain is self-contained with respect to circularity. The absence of ground-truth annotations for the three special test images and the lack of statistical robustness are correctness-risk issues and should be raised under a different review category, not as circular reasoning.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities or fitted model parameters. Its central claim depends on the representativeness of tiny hand-picked test data, the stability of metrics on a 19-image test set, and the fairness of unreported training settings.

free parameters (1)
  • confidence_threshold = 0.50
    Used in Tests 1-3 to decide which detections count as successful; chosen by hand and not justified or varied.
assumptions (4)
  • ad hoc to paper The three hand-picked test images are representative of not-so-common container damage.
    Tests 1-3 use three images chosen by the author; no predefined sampling criterion supports their representativeness.
  • domain assumption The 19-image test set yields a stable estimate of mAP, precision and recall.
    With 19 test images, a 4 percentage point mAP gap could easily be noise; no confidence intervals are reported.
  • ad hoc to paper All three models were trained under comparable conditions.
    No hyperparameters, epochs, image size, augmentation, or hardware details are reported, yet the comparison assumes fairness.
  • domain assumption The Roboflow-derived dataset annotations are correct and complete.
    Annotation quality and label completeness are not verified, which affects all reported metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Container damage detection using advanced computer vision model Yolov12 vs Yolov11 vs RF-DETR A comparative analysis." pith.science (2026). https://pith.science/paper/4XCNFQEO

@misc{pith2026250622517,
  author       = {Pith},
  title        = {Pith review of: Container damage detection using advanced computer vision model Yolov12 vs Yolov11 vs RF-DETR A comparative analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4XCNFQEO}},
  note         = {Machine review of arXiv:2506.22517}
}
read the original abstract

Containers are an integral part of the logistics industry and act as a barrier for cargo. A typical service life for a container is more than 20 years. However, overtime containers suffer various types of damage due to the mechanical as well as natural factors. A damaged container is a safety hazard for the employees handling it and a liability for the logistic company. Therefore, a timely inspection and detection of the damaged container is a key for prolonging service life as well as avoiding safety hazards. In this paper, we will compare the performance of the damage detection by three state-of-the-art advanced computer vision models Yolov12, Yolov11 and RF-DETR. We will use a dataset of 278 annotated images to train, validate and test the model. We will compare the mAP and precision of the model. The objective of this paper is to identify the model that is best suited for container damage detection. The result is mixed. mAP@50 score of Yolov11 and 12 was 81.9% compared to RF-DETR, which was 77.7%. However, while testing the model for not-so-common damaged containers, the RF-DETR model outperformed the others overall, exhibiting superiority to accurately detecting both damaged containers as well as damage occurrences with high confidence.

Figures

Figures reproduced from arXiv: 2506.22517 by the authors.

Figure 1
Figure 1. Type of container damage [14] [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. illustrates different sections of containers. A standard intermodal shipping container is a marvel of modular engineering, designed for durability and efficient cargo handling. Some key structural sections are front and rear end walls, side walls, roof panel, floor assembly, door assembly, corner castings and posts [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Yolov11 Benchmarking [3] 2.2 Key Features of Yolov12 Yolov12 introduced an attention-centric architecture that is totally different from traditional CNN-based architecture [5] in February 2025. Two major improvement factors for Yolov12 – Area Attention and Residual Efficient Layer Aggregation Networks (R-ELAN). Area Attention: Unlike traditional attention mechanisms that attend to individual elements such as a word … view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: One Dimensional area attention [7] [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Two-Dimensional area attention [7] Area attention allows models to focus on groups of adjacent elements as seen in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Benchmarking Yolov12 compared to previous versions: mAP vs. Latency and mAP vs. FLOPs [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Benchmarking RF-DETR compared to D-FINE, LW-DETR Yolov8, Yolov11 [8] [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Images used for the dataset 4 COMMON METRICS Here are the metrics that we are going to use to assess the model’s performance. 4.1 Precision In object detection, precision measures how accurate your model’s positive predictions are, essentially, how many of the objects …
Figure 9
Figure 9. Figure 9: How mAP changes as the model trains over time (Epochs) – Yolov11, Yolov12 and RF-DETR [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Fire damaged container for Test1 – Yolov11, Yolov12 and RF-DETR [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: Fire damaged container for Test2 – Yolov11, Yolov12 and RF-DETR [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: Bend container for Test 3 – Yolov11, Yolov12 and RF-DETR model [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 16 canonical work pages

  1. [1]

    Yolov11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, "Yolov11: An overview of the key architectural enhancements," arXiv preprint arXiv:2410.17725, 2024

    arXiv 2024

  2. [2]

    Yolo evolution: A comprehensive benchmark and architectural review of yolov12, yolo11, and their previous versions,

    N. Jegham, C. Y. Koh, M. Abdelatti, and A. Hendawi,"Yolo evolution: A comprehensive benchmark and architectural review of yolov12, yolo11, and their previous versions," 2024

    work page 2024

  3. [3]

    Ultralytics yolov11.https://docs.ultralytics.com/models/yolo11/s, 2024

    work page 2024

  4. [4]

    Robicheaux, M

    P. Robicheaux, M. Popov, A. Madan, I. Robinson, J. Nelson, D. Ramanan, N. Peri, Roboflow100-vl: A multi-domain object detection benchmark for vision-language models, Roboflow (2025)

    work page 2025

  5. [5]

    Two major improvement factors for Yolov12 – Area Attention and Residual Efficient Layer Aggregation Networks (R-ELAN)

    in February 2025. Two major improvement factors for Yolov12 – Area Attention and Residual Efficient Layer Aggregation Networks (R-ELAN). Area Attention: Unlike traditional attention mechanisms that attend to individual elements such as a word in a sentence (one-dimensional) or pixels in an image (two-dimensional). An area is a collection of structurally a...

    work page 2025

  6. [6]

    Y. Tian, Q. Ye, and D. Doermann,YOLOv12: Attention-centric real-time object detectors, 2025. [Online]. Available: https://github.com/sunsmarterjie/yolov12

    work page 2025

  7. [7]

    Mikomel, Area Attention [Online]

    M. Mikomel, Area Attention [Online]. Available: https://github.com/mikomel/area- attention/blob/main/README.md. . [Accessed: Jun. 8, 2025]

    work page 2025

  8. [8]

    Area attention,

    Y. Li, L. Kaiser, S. Bengio, and S. Si, "Area attention," International Conference on Machine Learning, May 2019, pp. 3846–3855, PMLR

    work page 2019

Show all 18 references
  1. [9]

    Sapkota, R

    R. Sapkota, R. H. Cheppally, A. Sharda, and M. Karkee, "RF-DETR Object Detection vs YOLOv12: A Study of Transformer-based and CNN-based Architectures for Single-Class and Multi-Class Greenfruit Detection in Complex Orchard Environments Under Label Ambiguity," arXiv preprint ar...

    2025 arXiv

  2. [10]

    Comparative Analysis of Deep Learning Models for Honeybee and Threat Detection at Hive Entrances,

    F. Ajam, "Comparative Analysis of Deep Learning Models for Honeybee and Threat Detection at Hive Entrances," 2025

    2025

  3. [11]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, "End-to-end object detection with transformers," in Proc. Eur. Conf. Comput. Vis. (ECCV), Cham, Switzerland: Springer, Aug. 2020, pp. 213–229

    2020

  4. [12]

    Automating container damage detection with the YOLO-NAS deep learning model,

    T. Nguyen Thi Phuong, G. S. Cho, and I. Chatterjee, “Automating container damage detection with the YOLO-NAS deep learning model,” Science Progress, vol. 108, no. 1, pp. 1–11, 2025

    2025

  5. [13]

    Detection of waste containers using computer vision,

    M. Valente, H. Silva, J. M. Caldeira, V. N. Soares, and P. D. Gaspar, “Detection of waste containers using computer vision,” Applied System Innovation, vol. 2, no. 1, p. 11, 2019

    2019

  6. [14]

    Automatic damage-detecting system for port container gate based on AI,

    C. Tang, P. Chen, and Y. Li, “Automatic damage-detecting system for port container gate based on AI,” inProc. 2020 9th Int. Conf. Comput. Pattern Recognit., Oct. 2020, pp. 146–151

    2020

  7. [15]

    Development of the container damage inspection system

    Oh JH, Hong SW, Choi GJ, et al. Development of the container damage inspection system. J Korean Soc Precis Eng 2005; 22: 82–88

    2005

  8. [16]

    Exploring the potential of climate-adaptive container building design under future climates scenarios in three different climate zones,

    J. Shen, B. Copertaro, X. Zhang, J. Koke, P. Kaufmann, and S. Krause, “Exploring the potential of climate-adaptive container building design under future climates scenarios in three different climate zones,”Sustainability, vol. 12, no. 1, p. 108, 2019. 12

    2019

  9. [17]

    Xueqi L, Qing L, Jinbo W,. et al. Container damage identification based on Fmask-RCNN. Commun Comp Inf Sci 2020; 1265: 12–22

    2020

  10. [18]

    Understanding of object detection based on CNN family and YOLO

    Juan D. Understanding of object detection based on CNN family and YOLO. J Phys 2018; 1004: 1–8

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.