Pith. sign in

REVIEW 3 major objections 2 minor 36 references

Inter-Class Relational Loss for Small Object Detection: A Case Study on License Plates

T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper proposes an inter-class relational loss that penalizes a predicted license-plate box when it falls outside its car box, and reports mAP50 gains of 10.3% and 1.6% on two detectors with no extra tuning.

desk verdict Plausible plug-in loss for small objects, but the abstract doesn't let you check the 10.3% claim; watch the plate-to-car association rule. read the letter →

arxiv 2508.14343 v1 pith:KJLPTOOZ submitted 2025-08-20 cs.CV cs.AI

classification cs.CVcs.AI
keywords smallobjectdetectioninter-classrelationallosslicenseplateIoU-basedYOLOv12UAV-DETRcontainmentpriormulti-license-platedataset
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 argues that IoU-based losses give tiny, flat gradients for small objects, so small objects learn too slowly when multiple objects are updated together. To fix this, the authors introduce an inter-class relational (ICR) loss based on a simple physical regularity: a license plate is attached to a car, so its predicted box should lie inside the car's box. When a predicted plate box is not contained in its car box, the loss adds a penalty that grows as the overlap area shrinks, giving the small object a gradient signal even when its IoU with the ground truth is near zero. The authors show the penalty can be bolted onto standard IoU losses, improving mAP50 by 10.3 points for YOLOv12-T and 1.6 points for UAV-DETR on a new license-plate dataset (SVMLP) with no extra tuning. The paper also releases SVMLP, a multi-license-plate benchmark with diverse real-world scenes.

What carries the argument

The central mechanism is the inter-class relational (ICR) loss penalty. It uses the annotated or predicted parent box (the car) as a spatial prior for the child box (the plate): the penalty is activated when the predicted child box is not contained within the parent box, and its magnitude is inversely proportional to the overlap area between the child prediction and the parent box. This makes the gradient for a small object non-negligible even when its IoU with the ground truth is near zero, because the overlap with the parent box provides a continuous error scale.

What would settle it

Train YOLOv12-T with and without the ICR loss on a dataset of small objects that are not contained in a larger parent box (e.g., small aerial-view vehicles or isolated small objects) and measure small-object mAP; if the ICR term does not improve or hurts performance, the containment signal—not merely an extra loss term—is what drives the reported gains. Alternatively, log gradient norms for small-object predictions during a training run; the ICR term must visibly increase those norms for the claimed mechanism to hold.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a small object's learning can be guided by its spatial containment in a larger, easier-to-detect object. Concretely, the ICR loss computes the overlap between the predicted plate box and the predicted car box, and when the plate box is not fully inside the car box, it adds a penalty inversely proportional to that overlap area. The penalty is architecture-agnostic: it is appended to the existing IoU-based loss, and the reported experiments show large mAP gains on a transformer detector and a CNN detector. The implication for the field is that small-object detection need not rely solely on sharper IoU formulations; exploiting cross-class

Load-bearing premise

The load-bearing premise is that a license plate's predicted bounding box should always be inside its car's bounding box and that the degree of overlap between the two boxes is a reliable, noise-free teaching signal; if scenes, camera angles, or occlusions break this containment relation, the penalty injects wrong gradients into training.

Editorial extensions

If this is right

  • Adding the ICR penalty to existing IoU-based losses produces measurable mAP gains with no extra hyperparameter tuning, at least for the two detectors tested.
  • The loss gives small objects a non-flat gradient signal early in training, which should accelerate convergence when multiple objects of different sizes are learned jointly.
  • The SVMLP dataset supplies a new multi-license-plate benchmark for future small-object detection work, with real-world diversity.
  • The containment prior is a general template: any small object attached to a larger parent (face on person, text on document) could receive the same treatment.
  • Because the penalty is additive, it can be dropped into existing training pipelines without changing the detector architecture or inference time.

Reading between the lines

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

  • The loss implicitly trusts the quality of the parent-box prediction: if the car detector misses or over-segments the car, the ICR penalty could push plate predictions in the wrong direction; this dependency is not isolated in the paper's ablations.
  • The inverse-overlap weighting resembles a softened containment constraint, and one could test whether the loss is formally equivalent to bounding the child box by the parent box under a reciprocal-overlap transformation; that equivalence would explain why it helps exactly in the near-zero-IoU regime.
  • The large gain gap between YOLOv12-T and UAV-DETR (10.3 vs 1.6 points) suggests the loss's benefit depends on the baseline's gradient behavior; applying the loss to more architectures would reveal whether the mechanism is universal or detector-specific.
  • A natural extension would be to apply the same containment loss to other physical parent-child pairs (e.g., a phone in a hand, a face occluding a body), but those relations are less rigid than plate-in-car, so the inverse-overlap weighting may need to be relaxed.
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

3 major / 2 minor

Summary. The paper proposes an Inter-Class Relational (ICR) loss for small object detection, demonstrated on license plate detection. The loss adds a penalty when a predicted license plate bounding box is not contained within its associated car bounding box, with the penalty inversely proportional to the overlap area. The authors claim that this relational penalty improves small-object gradient updates and report mAP50 gains of 10.3% on YOLOv12-T and 1.6% on UAV-DETR on a new SVMLP dataset, without additional hyperparameter tuning. The paper also introduces SVMLP as a new multi-license-plate benchmark.

Significance. If the claims are correct, ICR loss would be a simple, plug-and-play addition to existing IoU-based losses that leverages the reliable spatial parent–child relationship between cars and license plates. The proposed dataset could also be a useful resource for small-object detection in traffic scenes. However, the current manuscript (as provided) contains only the abstract; no equations, training configurations, association rules, full experimental comparisons, or statistical analyses are available. The central claim therefore remains unverified, and the unspecified plate-to-car association rule is a serious obstacle to reproducibility and to assessing whether the reported gains are due to the loss itself rather than to the matching mechanism.

major comments (3)
  1. [Abstract] The penalty condition 'When the predicted car plate's bounding box is not within its car' presupposes a definite association between each predicted plate and a specific car. The manuscript gives no rule for establishing this association. In a multi-object detector, both plate and car boxes are noisy outputs; a plate prediction may overlap zero, one, or several predicted car boxes, and the 'correct' car may not be detected. If the association uses predicted boxes, the resulting gradient depends on a non-differentiable matching step and can attach plate updates to the wrong vehicle. If it uses ground-truth parent annotations, the loss requires annotations that are not mentioned and may not generalize to datasets without such labels. This is load-bearing for the plug-and-play claim and must be specified, with a strategy for ambiguous cases (boundary plates, occlusions, absent cars).
  2. [Abstract / Results] The reported improvements of 10.3% and 1.6% mAP50 are presented without error bars, number of runs, or significance tests. The abstract also lacks any comparison to existing small-object loss functions (e.g., Wise-IoU, NWD, or other relational approaches), so it is unclear whether the gain is specific to ICR loss or could be obtained by tuning an existing loss. Additionally, no ablation separates the effect of the ICR penalty from the underlying IoU loss. The claim 'without any additional hyperparameter tuning' is unsubstantiated without a clear description of the baseline training protocol and the range of hyperparameters considered. These details are necessary to assess the magnitude and reliability of the improvement.
  3. [Abstract / Motivation] The motivation that IoU-based losses have an 'extremely flat gradient' for small objects is asserted without derivation or citation. The proposed inverse-overlap penalty also creates a sharp transition at the containment boundary: a small displacement from 'inside' to 'outside' changes the penalty abruptly, which could reintroduce the gradient instability that IoU losses are designed to avoid. The paper should provide a mathematical description of the loss, its gradient, and a small analysis of its behavior near the containment boundary, especially for plates that are only partially inside the car box due to annotation noise or perspective effects.
minor comments (2)
  1. [Abstract] The abstract states that code and dataset 'will be available soon.' If the dataset is essential to reproducing the results, the authors should clarify the annotation format, especially whether each plate box includes a parent-car identifier or an association attribute.
  2. [General] The title and abstract emphasize 'Inter-Class Relational Loss' but the loss is specifically a parent–child containment relation. The generality of the method to other inter-class relations (e.g., wheel on road, light on vehicle) is asserted but not demonstrated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ICR loss is a handcrafted penalty term and the reported mAP gains are empirical measurements, not predictions derived from fitted constants.

full rationale

The paper's central claim is empirical: adding the inter-class relational (ICR) loss penalty to existing IoU-based losses improves mAP50 by 10.3% and 1.6% on YOLOv12-T and UAV-DETR. The penalty is defined in the abstract: 'When the predicted car plate's bounding box is not within its car, a loss punishment is added to guide the learning, which is inversely proportional to the overlapped area of the car's and predicted car plate's bounding box.' This is a handcrafted training objective, not a quantity derived from the reported mAP values or fitted to the test set. There is no equation in which the output variable is defined as the input variable, no parameter fitted to a subset of data and then presented as a prediction on that same data, and no load-bearing self-citation—the abstract cites no prior work. The 'flat gradient' motivation is an asserted heuristic rather than a theorem derived from the loss, but an underived motivation is not circularity. The skeptic's concern about how 'its car' is assigned to a predicted plate is a substantive correctness and reproducibility question, not a circularity issue: it does not make the reported gains equivalent to the loss definition by construction. No appended limitation or self-referential passage is present in the provided text. Therefore no circular step is identified.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No explicit free parameters are disclosed. The loss formula's scaling is described only as 'inversely proportional to the overlapped area', which may imply an implicit multiplier not stated. The assumptions listed are all drawn from the abstract and are plausible, but unverified.

assumptions (3)
  • domain assumption A license plate is spatially attached to its car and normally lies within the car's bounding box.
    Core premise of the ICR loss; referenced in abstract as 'a car plate is attached to a car in a similar position'.
  • domain assumption IoU-based losses have extremely flat gradients for small objects, which harms their training update.
    The motivation for the new loss, stated in the abstract but not derived or empirically established in the abstract.
  • domain assumption The added relational penalty does not degrade learning of other objects.
    The abstract claims the method works 'while not sacrificing the learning efficiency of other objects', but this is asserted as a design goal, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inter-Class Relational Loss for Small Object Detection: A Case Study on License Plates." pith.science (2026). https://pith.science/paper/KJLPTOOZ

@misc{pith2026250814343,
  author       = {Pith},
  title        = {Pith review of: Inter-Class Relational Loss for Small Object Detection: A Case Study on License Plates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KJLPTOOZ}},
  note         = {Machine review of arXiv:2508.14343}
}
abstract

In one-stage multi-object detection tasks, various intersection over union (IoU)-based solutions aim at smooth and stable convergence near the targets during training. However, IoU-based losses fail to correctly update the gradient of small objects due to an extremely flat gradient. During the update of multiple objects, the learning of small objects' gradients suffers more because of insufficient gradient updates. Therefore, we propose an inter-class relational loss to efficiently update the gradient of small objects while not sacrificing the learning efficiency of other objects based on the simple fact that an object has a spatial relationship to another object (e.g., a car plate is attached to a car in a similar position). When the predicted car plate's bounding box is not within its car, a loss punishment is added to guide the learning, which is inversely proportional to the overlapped area of the car's and predicted car plate's bounding box. By leveraging the spatial relationship at the inter-class level, the loss guides small object predictions using larger objects and enhances latent information in deeper feature maps. In this paper, we present twofold contributions using license plate detection as a case study: (1) a new small vehicle multi-license plate dataset (SVMLP), featuring diverse real-world scenarios with high-quality annotations; and (2) a novel inter-class relational loss function designed to promote effective detection performance. We highlight the proposed ICR loss penalty can be easily added to existing IoU-based losses and enhance the performance. These contributions improve the standard mean Average Precision (mAP) metric, achieving gains of 10.3% and 1.6% in mAP$^{\text{test}}_{50}$ for YOLOv12-T and UAV-DETR, respectively, without any additional hyperparameter tuning. Code and dataset will be available soon.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages

  1. [1]

    C. Cao, B. Wang, W. Zhang, X. Zeng, X. Yan, Z. Feng, Y. Liu, and Z. Wu. An improved faster R-CNN for small object detection. IEEE Access , 7:106838--106846, 2019

  2. [2]

    Cheng, X

    G. Cheng, X. Yuan, X. Yao, K. Yan, Q. Zeng, X. Xie, and J. Han. Towards large-scale small object detection: Survey and benchmarks. IEEE Trans. Pattern Anal. Mach. Intell. , 45(11):13467--13488, 2023

  3. [3]

    Dubey, F

    S. Dubey, F. Olimov, M. A. Rafique, and M. Jeon. Improving small objects detection using transformer. J. Vis. Commun. Image Represent. , 89:103620, 2022

  4. [4]

    J. C. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for online learning and stochastic optimization. J. Mach. Learn. Res. , 12:2121--2159, 2011

  5. [5]

    Y. Gong, L. Deng, S. Tao, X. Lu, P. Wu, Z. Xie, Z. Ma, and M. Xie. Unified chinese license plate detection and recognition with high efficiency. J. Vis. Commun. Image Represent. , 86:103541, 2022

  6. [6]

    C. Han, G. Gao, and Y. Zhang. Real-time small traffic sign detection with revised faster-rcnn. Multim. Tools Appl. , 78(10):13263--13278, 2019

  7. [7]

    Huang, H

    Y. Huang, H. Liu, H. Shuai, and W. Cheng. DQ-DETR: DETR with dynamic query for tiny object detection. In A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, editors, Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXXVI , volume 15134 of Lecture Notes in Computer ...

  8. [8]

    Ismail, M

    A. Ismail, M. Mehri, A. Sahbani, and N. E. B. Amara. ELP 1.0: A comprehensive and geographically diverse dataset of european license plates. IEEE Trans. Intell. Transp. Syst. , 26(6):7573--7586, 2025

Show all 36 references
  1. [9]

    S. Ji, Q. Ling, and F. Han. An improved algorithm for small object detection based on YOLO v4 and multi-scale contextual information. Comput. Electr. Eng. , 105:108490, 2023

  2. [10]

    N. S. Keskar, D. Mudigere, J. Nocedal, M. Smelyanskiy, and P. T. P. Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track...

  3. [11]

    Khanam and M

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

  4. [12]

    Laroca, E

    R. Laroca, E. Severo, L. A. Zanlorensi, L. S. Oliveira, G. R. Gon c alves, W. R. Schwartz, and D. Menotti. A robust real-time automatic license plate recognition based on the YOLO detector. In 2018 International Joint Conference on Neural Networks, IJCNN 2018, Rio de Janeiro, ...

  5. [13]

    T. Lin, P. Goyal, R. B. Girshick, K. He, and P. Doll \' a r. Focal loss for dense object detection. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017 , pages 2999--3007. IEEE Computer Society, 2017

  6. [14]

    T. Lin, M. Maire, S. J. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll \' a r, and C. L. Zitnick. Microsoft COCO: common objects in context. In D. J. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, editors, Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Sw...

  7. [15]

    W. Lv, Y. Zhao, Q. Chang, K. Huang, G. Wang, and Y. Liu. Rt-detrv2: Improved baseline with bag-of-freebies for real-time detection transformer. arXiv preprint , arXiv:2407.17140, 2024

  8. [16]

    Q. Ming, L. Miao, Z. Ma, L. Zhao, Z. Zhou, X. Huang, Y. Chen, and Y. Guo. Deep dive into gradients: Better optimization for 3d object detection with gradient-corrected iou supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC,...

  9. [17]

    Ouyang, X

    W. Ouyang, X. Wang, C. Zhang, and X. Yang. Factors in finetuning deep model for object detection with long-tail distribution. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016 , pages 864--873. IEEE Computer Soc...

  10. [18]

    Z. Peng, Y. Gao, S. Mu, and S. Xu. Toward reliable license plate detection in varied contexts: Overcoming the issue of undersized plate annotations. IEEE Trans. Intell. Transp. Syst. , 25(11):18107--18121, 2024

  11. [19]

    S. H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. D. Reid, and S. Savarese. Generalized intersection over union: A metric and A loss for bounding box regression. arXiv preprint , arXiv:1902.09630, 2019

  12. [20]

    N. P. Sutramiani, N. Suciati, and D. Siahaan. MAT-AGCA: multi augmentation technique on small dataset for balinese character recognition using convolutional neural network. ICT Express , 7(4):521--529, 2021

  13. [21]

    Y. Tian, Q. Ye, and D. S. Doermann. Yolov12: Attention-centric real-time object detectors. arXiv preprint , arXiv:2502.12524, 2025

  14. [22]

    C. Wang, I. Yeh, and H. M. Liao. Yolov9: Learning what you want to learn using programmable gradient information. In A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, editors, Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, Septem...

  15. [23]

    H. Wang, C. Liu, Y. Cai, L. Chen, and Y. Li. Yolov8-qsd: An improved small object detection algorithm for autonomous vehicles based on yolov8. IEEE Trans. Instrum. Meas. , 73:1--16, 2024

  16. [24]

    J. Wang, C. Xu, W. Yang, and L. Yu. A normalized gaussian wasserstein distance for tiny object detection. arXiv preprint , arXiv:2110.13389, 2021

  17. [25]

    Q. Wang, X. Lu, C. Zhang, Y. Yuan, and X. Li. LSV-LP: large-scale video-based license plate detection and recognition. IEEE Trans. Pattern Anal. Mach. Intell. , 45(1):752--767, 2023

  18. [26]

    S. Wu, J. Yang, X. Wang, and X. Li. Iou-balanced loss functions for single-stage object detection. Pattern Recognit. Lett. , 156:96--103, 2022

  19. [27]

    Xiuling, W

    Z. Xiuling, W. Huijuan, S. Yu, C. Gang, Z. Suhua, and Y. Quanbo. Starting from the structure: A review of small object detection based on deep learning. Image Vis. Comput. , 146:105054, 2024

  20. [28]

    X. Yu, Y. Gong, N. Jiang, Q. Ye, and Z. Han. Scale match for tiny person detection. In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020 , pages 1246--1254. IEEE , 2020

  21. [29]

    X. Yuan, G. Cheng, K. Yan, Q. Zeng, and J. Han. Small object detection via coarse-to-fine proposal generation and imitation learning. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023 , pages 6294--6304. IEEE , 2023

  22. [30]

    Zhang, K

    H. Zhang, K. Liu, Z. Gan, and G. Zhu. UAV-DETR: efficient end-to-end object detection for unmanned aerial vehicle imagery. arXiv preprint , arXiv:2501.01855, 2025

  23. [31]

    Zhang and S

    H. Zhang and S. Zhang. Focaler-iou: More focused intersection over union loss. arXiv preprint , arXiv:2401.10525, 2024

  24. [32]

    Zhang, M

    Y. Zhang, M. Ye, G. Zhu, Y. Liu, P. Guo, and J. Yan. FFCA-YOLO for small object detection in remote sensing images. IEEE Trans. Geosci. Remote. Sens. , 62:1--15, 2024

  25. [33]

    Y. Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y. Liu, and J. Chen. Detrs beat yolos on real-time object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 , pages 16965--16974. IEEE , 2024

  26. [34]

    Zheng, P

    Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, and D. Ren. Distance-iou loss: Faster and better learning for bounding box regression. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conf...

  27. [35]

    Zheng, P

    Z. Zheng, P. Wang, D. Ren, W. Liu, R. Ye, Q. Hu, and W. Zuo. Enhancing geometric factors in model learning and inference for object detection and instance segmentation. IEEE Trans. Cybern. , 52(8):8574--8586, 2022

  28. [36]

    D. Zhou, J. Fang, X. Song, C. Guan, J. Yin, Y. Dai, and R. Yang. Iou loss for 2d/3d object detection. arXiv preprint , arXiv:1908.03851, 2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.