Pith. sign in

REVIEW 5 major objections 5 minor 39 references

Pseudo-Labeling Driven Refinement of Benchmark Object Detection Datasets via Analysis of Learning Patterns

T0 review · 5 major / 5 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 Annotation errors in MS-COCO can be found and fixed automatically via loss-gradient anomaly detection and pseudo-labeling, yielding MJ-COCO, which improves detectors on clean benchmarks.

desk verdict MJ-COCO is a real new dataset, but the paper's own Tables 8–12 contradict its headline claim of consistent improvement over MS-COCO. read the letter →

arxiv 2506.00997 v1 pith:GJMEONP7 submitted 2025-06-01 cs.CV

Min Je Kim , Muhammad Munsif , Altaf Hussain , Hikmat Yar , Sung Wook Baik This is my paper
classification cs.CV
keywords objectdetectiondatasetannotationrefinementpseudo-labelinganomalylossandgradientmonitoringMS-COCOMJ-COCOsmall
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

MS-COCO, the standard object-detection benchmark, carries many wrong annotations: the paper's own human expert judged 44.9% of its 117,267 images to be incorrectly labeled. The paper's claim is that these errors can be found automatically from training statistics and fixed at scale by a four-stage pseudo-labeling pipeline, yielding a corrected dataset called MJ-COCO. Trained detectors then generalize better: across one-stage and two-stage models, MJ-COCO-trained weights improve Average Precision and small-object AP on the cleaner Sama COCO and Objects365 validation sets, and small-object annotations grow by more than 200,000. If true, the framework offers a practical way to repair benchmark datasets without re-labeling every image by hand.

What carries the argument

The load-bearing mechanism is a Faster R-CNN training monitor that feeds an 8-dimensional vector of losses and gradients—RPN classification/regression and ROI classification/regression—into a linear AutoEncoder; reconstruction error is the anomaly score. The regression loss is normalized by $L' = L/(N_{\text{matching}})\cdot(1+N_{\text{FP}})$, and zero-loss images with ground-truth objects are assigned the maximum historical loss so detection failures do not vanish. The second mechanism is the four-stage pseudo-labeling pipeline: six invertible image transformations give candidate boxes, boxes with IoU at least 0.8 merge by confidence, a ResNet-50 classifier validates low-confidence boxes through confidence-tiered rules, and Grad-CAM variance and concentration adjust or reject boxes. Together these mechanisms convert hard-to-learn images into corrected annotations.

What would settle it

Have a random sample of MJ-COCO images independently re-annotated by several human annotators under a written protocol and compare their correct/incorrect calls with the Oracle's and their boxes with MJ-COCO's pseudo-labels; if the Oracle's 44.9% error rate is not reproduced or the pseudo-labels disagree with the annotators on a large fraction of boxes, the central claim is not supported.

Watch

Extended reading notes

Core claim

The central discovery is that the learning pattern of a two-stage detector—its normalized regression loss and gradient magnitudes on each image—separates bad annotations from good ones. A linear AutoEncoder over the eight monitored signals flags anomalous images, and the paper reports that normalization plus loss weighting raises error-detection AUROC from 0.8003 to 0.8436 and F1 from 0.6406 to 0.7016. The flagged images are not discarded but re-annotated through six-view invertible transformations, IoU-based duplicate merging, confidence-tiered class verification by a ResNet-50 classifier, and Grad-CAM alignment. The resulting MJ-COCO has substantially more small-object annotations and is claimed to train detectors that consistently outperform MS-COCO-trained ones on high-quality external validation sets, with especially clear gains on Sama COCO and Objects365.

Load-bearing premise

The load-bearing premise is that the Oracle's binary judgment—one unnamed human expert's correct/incorrect call on all 117,267 MS-COCO images—is reliable ground truth, even though the paper gives no details on who that expert was, what instructions they followed, or whether other annotators would agree.

Editorial extensions

If this is right

  • Detectors trained on MJ-COCO outperform MS-COCO-trained counterparts on high-quality external validation sets, with notable AP and AP_S gains on Sama COCO and Objects365.
  • MJ-COCO adds more than 200,000 small-object annotations, improving coverage for classes such as Apple, Carrot, Book, and Bottle.
  • The two-stage detection monitor plus linear AutoEncoder detects annotation errors better than the GAN, diffusion, and Transformer-AutoEncoder baselines tested, especially when loss and gradient features are normalized.
  • The pseudo-labeling approach is scalable to large datasets without per-image manual correction.
  • MJ-COCO is publicly released, so the claimed gains can be independently replicated.

Reading between the lines

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

  • The reported gains are concentrated on clean external validation sets; on the original MS-COCO validation set, MJ-COCO-trained models often show slightly lower AP (for example YOLOX drops from 0.453 to 0.433). A likely consequence is that the benefit is real for benchmark users who evaluate against Sama-style clean labels, but would not show up under the old COCO evaluation protocol.
  • Because the Oracle's binary judgment is the ground truth for both the reported error rate and the AutoEncoder evaluation, the method's measured accuracy would probably shift if the Oracle were replaced by a panel; the paper gives no way to estimate that shift.
  • The expert objects recognizer is described only as a separately trained ResNet-50 classifier, so the Stage-3 confidence thresholds carry an undocumented dependence on that classifier's calibration; a testable extension would be to measure how sensitive MJ-COCO's label set is to swapping that classifier.
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

5 major / 5 minor

Summary. The paper proposes a three-stage framework for refining the MS-COCO object detection dataset: (1) an anomaly-detection stage that uses Faster R-CNN loss and gradient statistics fed into a linear AutoEncoder to flag images with annotation errors; (2) a four-stage pseudo-labeling pipeline that generates, filters, class-verifies, and spatially adjusts boxes using invertible transformations, IoU merging, a ResNet-50 classifier, and Grad-CAM; (3) an evaluation of the resulting 'MJ-COCO' dataset by training five one- and two-stage detectors and testing on MS-COCO, Sama-COCO, Objects365, and PASCAL validation sets. The abstract claims that models trained on MJ-COCO consistently outperform those trained on MS-COCO, especially on high-quality validation sets, and that MJ-COCO adds more than 200,000 small-object annotations.

Significance. If the central claim held, the work would be a valuable contribution: a scalable, automatic way to refine the most widely used object detection benchmark, with the resulting dataset released publicly. The pipeline is broad and the paper evaluates several detector architectures, which is a strength. However, the paper's own cross-validation results directly contradict the headline claim on the canonical MS-COCO validation set, the small gains on Sama-COCO and Objects365 are not statistically supported, and the evaluation design has a circularity concern because the pseudo-labels are generated by the same family of detectors used in the evaluation. The potential significance is high, but the current evidence is insufficient to support the stated conclusions.

major comments (5)
  1. [Section IV-F, Tables 8–12] The abstract's claim that models trained on MJ-COCO 'consistently outperformed those trained on MS-COCO' is directly contradicted by the paper's own cross-validation on the original MS-COCO validation set, where all five detectors score lower when trained on MJ-COCO: RetinaNet AP 0.349 vs 0.367, YOLOv3 0.286 vs 0.294, YOLOX 0.433 vs 0.453, Faster R-CNN 0.348 vs 0.354, and Libra R-CNN 0.355 vs 0.360. Since MS-COCO val is the canonical benchmark for the dataset being refined, a training set that loses roughly 1–2 AP there does not support the unqualified 'consistently outperformed' claim; the conclusion in Section V softens to 'outperformed or matched,' which is inconsistent with the abstract.
  2. [Section III-A, Figure 2] The Oracle, an unnamed human expert, is used as ground truth for both evaluating the autoencoder error detector (Section IV-D) and selecting errored images that enter pseudo-labeling (Section IV-E). The paper provides no information about the Oracle's identity, annotation instructions, number of annotators, inter-annotator agreement, or quality control. The reported 44.9% error rate and all downstream pseudo-labels rely on this binary judgment; without a reproducible description of the Oracle labeling process, the foundational ground truth of the entire pipeline is unverifiable.
  3. [Section III-B and IV-F] The pseudo-labels are generated by detection models that belong to the same architecture family as the evaluated detectors: Figure 2 indicates CO-DETR for the transformation stage, and the error-detection stage uses Faster R-CNN. The improvements on Sama-COCO and Objects365 may therefore reflect the pseudo-label generator's inductive biases rather than objective label accuracy; the consistent degradation on MS-COCO val is consistent with this concern. To rule out circularity, the evaluation should include a detector family not used in pseudo-labeling, or the pseudo-labels should be produced from a different model archetype.
  4. [Section IV-F] The reported gains on Sama-COCO and Objects365 are small in absolute terms, for example AP improvements of +0.002 to +0.014 in Tables 8–12 for the positive cases, with some metrics changing by only 0.001. No repeated runs, random seeds, or confidence intervals are reported, so these differences are within typical training variance and the 'consistently outperformed' claim is not statistically established even for the clean benchmarks.
  5. [Algorithm 2, lines 11 and 15, and Section IV-E] The pseudo-labeling procedure contains internally inconsistent threshold specifications. Equation 9 and the surrounding prose disagree: Eq. 9 uses a condition of IoU < 0.6 with a max operator, while the prose states that boxes with IoU ≥ 0.8 are merged and later refers to a 'high IoU threshold of 0.8.' Similarly, the Stage 3 confidence thresholds in Algorithm 2 (sr ≥ 0.6, then 0.3 ≤ sr < 0.6 with pc conditions of 0.8 or 0.9) disagree with the prose in Section IV-E, which says that boxes with confidence between 0.3 and 0.6 require a classification probability of at least 0.7. These inconsistencies make the pseudo-labeling protocol irreproducible as written.
minor comments (5)
  1. [Section IV-B] The text says MS-COCO contains 'over 330,000 images,' but the analyzed set is stated as 117,267 images; the relationship between these numbers should be clarified.
  2. [Contributions bullet] The phrase 'Objects36' is a typo; it should read 'Objects365.'
  3. [Supplementary tables] Supplementary Tables 1–4 are referenced for the alpha, beta, gamma, delta ablations and class-wise AP results, but they are not provided with the submission; these results should be included for review.
  4. [Section IV-C, Eq. 15] Equation 15 defines AP as a simple mean over IoU thresholds, while the standard COCO AP is the mean over IoU thresholds and categories; this simplification should be stated explicitly.
  5. [Section IV-D] The reported breakdown of the 117,267 analyzed images gives TN+TP+FP+FN = 117,266, an off-by-one inconsistency that should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No construction-level circularity; the pseudo-labeling pipeline is tested on held-out human-annotated sets, though the abstract's 'consistently outperformed' claim is contradicted by the paper's own MS-COCO val results.

full rationale

I traced the claimed derivation chain stage by stage. Stage 1 (Oracle) is an external human judgment of correct vs incorrect annotations, not derived from the proposed AutoEncoder. Stage 2 fits the AutoEncoder and weighting hyperparameters (e.g., lambda, alpha-delta) to Oracle-labeled MS-COCO data, but these fits are not renamed as the downstream claim: the downstream claim is that detectors trained on MJ-COCO improve AP on held-out validation sets. Stage 3 generates pseudo-labels with an object detector (Figure 2 indicates CO-DETR) and a ResNet-50 image classifier, while Stage 4 evaluates with RetinaNet, YOLOv3, YOLOX, Faster R-CNN, and Libra R-CNN. The paper does not state that the pseudo-label generator is one of the five evaluated detectors, and the ResNet-50 classifier is not an AP-evaluated model. No equation in the paper defines downstream AP as a function of the anomaly scores or of the pseudo-labeling loss, and no fitted parameter is presented as a prediction of AP. The external validation sets (Sama, Objects365, PASCAL) provide independent human-annotated targets, and MS-COCO val provides a falsifying check. The main problem is not circularity but an internal evidence contradiction: Tables 8-12 show all five MJ-COCO-trained models lose AP on MS-COCO val (e.g., RetinaNet 0.349 vs 0.367, YOLOX 0.433 vs 0.453), so the abstract's unqualified 'consistently outperformed' is unsupported. That is a correctness/reporting issue, not a circular derivation. No load-bearing self-citations or imported uniqueness theorems appear in the reference list. Therefore no circular step meets the required evidentiary bar.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim rests on an unverified Oracle, on hand-tuned thresholds, and on the assumption that loss-gradient statistics identify label errors. Many parameters are chosen empirically, and unspecified Grad-CAM thresholds make the pipeline irreproducible.

free parameters (8)
  • loss-gradient weight lambda = 0.7 (or 0.9 for AUROC/PRAUC)
    Chosen empirically in Table 4 and supplementary; maximizes anomaly detection accuracy/F1.
  • feature-specific weights alpha,beta,gamma,delta = 0.3, 0.1, 0.3, 0.3
    Selected by ablation in supplementary Table 2; used in combined weighting.
  • anomaly detection cutoff = top 35%
    Fixed threshold used for anomaly flagging; no principled justification.
  • multi-view consistency count = at least 4 of 6 views
    Hand-chosen in Algorithm 2 line 5; controls pseudo-label acceptance.
  • IoU duplicate removal threshold = 0.8 (Eq. 9), 0.6 (Algorithm 2)
    Inconsistent across paper; hand-chosen, affects merging of bounding boxes.
  • confidence thresholds for classifier verification = 0.5/0.6, 0.3, 0.2, 0.1, 0.08 with Top-3/Top-1 matching and pc requirements
    Hierarchical hand-tuned rules in Algorithm 2 lines 14-17 and Section III-B; determine which low-confidence boxes are accepted.
  • Grad-CAM variance and concentration thresholds (alpha, beta) = unspecified
    Used in Algorithm 2 lines 22-25; values never reported, making refinement stage irreproducible.
  • number of transformations = 6
    Hand-chosen set of flips and scales in Section III-B.
assumptions (6)
  • domain assumption The Oracle's manual binary classification of all 117,267 MS-COCO images as correctly or erroneously annotated is accurate and unbiased.
    Used as ground truth for evaluating the anomaly detector and defining the pseudo-labeling input set (Section III-A). No quality control or inter-annotator agreement is reported.
  • domain assumption Loss and gradient statistics recorded during Faster R-CNN training reliably separate annotation errors from hard but correctly labeled examples.
    This is the core principle for the error detection stage (Section III-A). The paper provides no independent evidence that high loss/gradient implies label error rather than inherent difficulty.
  • ad hoc to paper Consistent detection across at least 4 of 6 image transformations indicates a real object with a reliable bounding box.
    The consistency count threshold is introduced for this pipeline (Algorithm 2 line 5) and is not derived from any theory or benchmark.
  • ad hoc to paper Grad-CAM activation maps with low variance and high concentration imply the current bounding box is correctly localized.
    The spatial adjustment stage (Algorithm 2 lines 20-27) assumes activation statistics measure box quality, with thresholds alpha and beta left unspecified.
  • domain assumption The ResNet-50 classifier trained on Objects365, PASCAL VOC, Caltech256, ImageNet, CIFAR, and Roboflow provides valid class-consistency verification after manual mapping to the MS-COCO taxonomy.
    Stage 3 of pseudo-labeling relies on this external classifier; manual class integration (Section IV-B) may introduce mapping errors.
  • domain assumption Detector performance on the Sama COCO and Objects365 validation sets is a valid proxy for training-label quality.
    The paper uses out-of-domain validation performance to argue MJ-COCO is better, but the standard MS-COCO validation set shows the opposite pattern, indicating the proxy is not universally valid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pseudo-Labeling Driven Refinement of Benchmark Object Detection Datasets via Analysis of Learning Patterns." pith.science (2026). https://pith.science/paper/GJMEONP7

@misc{pith2026250600997,
  author       = {Pith},
  title        = {Pith review of: Pseudo-Labeling Driven Refinement of Benchmark Object Detection Datasets via Analysis of Learning Patterns},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GJMEONP7}},
  note         = {Machine review of arXiv:2506.00997}
}
read the original abstract

Benchmark object detection (OD) datasets play a pivotal role in advancing computer vision applications such as autonomous driving, and surveillance, as well as in training and evaluating deep learning-based state-of-the-art detection models. Among them, MS-COCO has become a standard benchmark due to its diverse object categories and complex scenes. However, despite its wide adoption, MS-COCO suffers from various annotation issues, including missing labels, incorrect class assignments, inaccurate bounding boxes, duplicate labels, and group labeling inconsistencies. These errors not only hinder model training but also degrade the reliability and generalization of OD models. To address these challenges, we propose a comprehensive refinement framework and present MJ-COCO, a newly re-annotated version of MS-COCO. Our approach begins with loss and gradient-based error detection to identify potentially mislabeled or hard-to-learn samples. Next, we apply a four-stage pseudo-labeling refinement process: (1) bounding box generation using invertible transformations, (2) IoU-based duplicate removal and confidence merging, (3) class consistency verification via expert objects recognizer, and (4) spatial adjustment based on object region activation map analysis. This integrated pipeline enables scalable and accurate correction of annotation errors without manual re-labeling. Extensive experiments were conducted across four validation datasets: MS-COCO, Sama COCO, Objects365, and PASCAL VOC. Models trained on MJ-COCO consistently outperformed those trained on MS-COCO, achieving improvements in Average Precision (AP) and APS metrics. MJ-COCO also demonstrated significant gains in annotation coverage: for example, the number of small object annotations increased by more than 200,000 compared to MS-COCO.

Figures

Figures reproduced from arXiv: 2506.00997 by the authors.

Figure 1
Figure 1. Visual representation of annotation errors in the MS [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework: Consist of three stages: data preparation, error detection using a hybrid two [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of loss and gradient variations across training epochs. (a) illustrates the changes in the mean and median [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of AUROC and PRAUC curves based on varying thresholds. The blue curve indicates the results of anomaly [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Comparison of annotation errors in the MS-COCO dataset and corrections made by the proposed pseudo-labeling method. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Bounding box refinement results based on activation regions using Grad-CAM. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 30 canonical work pages

  1. [1]

    Tide: A general toolbox for identifying object detection errors,

    D. Bolya, S. Foley, J. Hays, and J. Hoffman, “Tide: A general toolbox for identifying object detection errors,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16 . Springer, 2020, pp. 558–573

    work page 2020

  2. [2]

    A trainable pedestrian detection system,

    C. Papageorgiou, T. Evgeniou, and T. Poggio, “A trainable pedestrian detection system,” in Proc. of Intelligent Vehicles , 1998, pp. 241–246

    work page 1998

  3. [3]

    Histograms of oriented gradients for human detection,

    N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR’05) , vol. 1. Ieee, 2005, pp. 886–893

    2005

  4. [4]

    The 2005 pascal visual object classes challenge,

    M. Everingham, A. Zisserman, C. K. Williams, L. Van Gool, M. Allan, C. M. Bishop, O. Chapelle, N. Dalal, T. Dese- laers, G. Dorkó et al. , “The 2005 pascal visual object classes challenge,” in Machine learning challenges. evaluating predic- tive uncertainty, visual object classification, and recognising tectual entailment: first PASCAL machine learning ch...

    work page 2005

  5. [5]

    The pascal visual object classes challenge 2012 (voc2012) development kit,

    M. Everingham and J. Winn, “The pascal visual object classes challenge 2012 (voc2012) development kit,” Pattern Analysis, Statistical Modelling and Computational Learning, Tech. Rep , vol. 8, no. 5, pp. 2–5, 2011

    work page 2012

  6. [6]

    MS COCO: Common Objects in Context,

    C. Consortium, “MS COCO: Common Objects in Context,” https://cocodataset.org/#download, 2014, accessed: 2025-04-07

    work page 2014

  7. [7]

    Bench- marking a benchmark: How reliable is ms-coco?

    E. Zimmermann, J. Szeto, J. Pasquero, and F. Ratle, “Bench- marking a benchmark: How reliable is ms-coco?” arXiv preprint arXiv:2311.02709, 2023

    arXiv 2023

  8. [8]

    The effect of improving annotation quality on object detection datasets: A preliminary study,

    J. Ma, Y . Ushiku, and M. Sagara, “The effect of improving annotation quality on object detection datasets: A preliminary study,” in Proceedings Of The IEEE/CVF conference on com- puter vision and pattern recognition , 2022, pp. 4850–4859

    work page 2022

Show all 39 references
  1. [9]

    The pascal visual object classes challenge 2006 (voc 2006) results (technical report). september 2006,

    M. Everingham, A. Zisserman, C. Williams, L. Van Gool et al., “The pascal visual object classes challenge 2006 (voc 2006) results (technical report). september 2006,” The PASCAL2006 dataset can be downloaded at http://www. pascal-network. org/challenges/VOC/voc2006, 2006

    2006

  2. [10]

    The pascal visual object classes challenge 2007 (voc 2007) results (2007),

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, “The pascal visual object classes challenge 2007 (voc 2007) results (2007),” 2008

    2007

  3. [11]

    Pascal voc 2008 challenge,

    D. Hoiem, S. K. Divvala, and J. H. Hays, “Pascal voc 2008 challenge,” World Literature Today , vol. 24, no. 1, pp. 1–4, 2009

    2008

  4. [12]

    The pascal visual object classes challenge 2009 (voc2009) results. http,

    M. Everingham, L. VanGool, C. Williams, J. Winn, and A. Zisserman, “The pascal visual object classes challenge 2009 (voc2009) results. http,” in www. pascal-network. org/challenges/VOC/voc2009/workshop/index. html, 2009

    2009

  5. [13]

    The pascal visual object classes (voc) chal- lenge,

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, “The pascal visual object classes (voc) chal- lenge,” International journal of computer vision , vol. 88, pp. 303–338, 2010

    2010

  6. [14]

    The pascal visual object classes challenge 2011 (voc2011) results, 2011< http://www. pascal-network. org/challenges,

    M. Everingham, L. Van Gool, C. Williams, J. Winn, and A. Zisserman, “The pascal visual object classes challenge 2011 (voc2011) results, 2011< http://www. pascal-network. org/challenges,” in VOC/voc2011/workshop/index. html

    2011

  7. [15]

    Imbalance problems in object detection: A review,

    K. Oksuz, B. C. Cam, S. Kalkan, and E. Akbas, “Imbalance problems in object detection: A review,” IEEE transactions on pattern analysis and machine intelligence , vol. 43, no. 10, pp. 3388–3415, 2020

    2020

  8. [16]

    Breaking beyond coco object detection

    A. Borji, “Breaking beyond coco object detection.”

  9. [17]

    Rethinking pascal-voc and ms-coco dataset for small object detection,

    K. Tong and Y . Wu, “Rethinking pascal-voc and ms-coco dataset for small object detection,” Journal of Visual Communication and Image Representation , vol. 93, p. 103830, 2023

    2023

  10. [18]

    Diagnosing state-of-the- art object proposal methods,

    H. Zhu, S. Lu, J. Cai, and Q. Lee, “Diagnosing state-of-the- art object proposal methods,” arXiv preprint arXiv:1507.04512, 2015

    2015 arXiv

  11. [19]

    Can we trust bounding box annotations for object detection?

    J. Murrugarra-Llerena, L. N. Kirsten, and C. R. Jung, “Can we trust bounding box annotations for object detection?” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4813–4822

    2022

  12. [20]

    Objectlab: Automated diagnosis of mislabeled images in object detection data,

    U. Tkachenko, A. Thyagarajan, and J. Mueller, “Objectlab: Automated diagnosis of mislabeled images in object detection data,” arXiv preprint arXiv:2309.00832 , 2023

    2023 arXiv

  13. [21]

    Identifying label errors in object detection datasets by loss inspection,

    M. Schubert, T. Riedlinger, K. Kahl, D. Kröll, S. Schoenen, S. Šegvi´c, and M. Rottmann, “Identifying label errors in object detection datasets by loss inspection,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 4582–4591

    2024

  14. [22]

    Training robust object detectors from noisy category labels and imprecise bounding boxes,

    Y . Xu, L. Zhu, Y . Yang, and F. Wu, “Training robust object detectors from noisy category labels and imprecise bounding boxes,” IEEE Transactions on Image Processing , vol. 30, pp. 5782–5792, 2021

    2021

  15. [23]

    Robust object detection with inaccurate bounding boxes,

    C. Liu, K. Wang, H. Lu, Z. Cao, and Z. Zhang, “Robust object detection with inaccurate bounding boxes,” in European Conference on Computer Vision . Springer, 2022, pp. 53–69

    2022

  16. [24]

    Pseco: Pseudo labeling and consistency training for semi-supervised object detection,

    G. Li, X. Li, Y . Wang, Y . Wu, D. Liang, and S. Zhang, “Pseco: Pseudo labeling and consistency training for semi-supervised object detection,” in European Conference on Computer Vision. Springer, 2022, pp. 457–472

    2022

  17. [25]

    Dense teacher: Dense pseudo-labels for semi-supervised ob- ject detection,

    H. Zhou, Z. Ge, S. Liu, W. Mao, Z. Li, H. Yu, and J. Sun, “Dense teacher: Dense pseudo-labels for semi-supervised ob- ject detection,” in European Conference on Computer Vision . Springer, 2022, pp. 35–50

    2022

  18. [26]

    Pseudo-label enhancement for weakly supervised object detection using self- supervised vision transformer,

    K. Yang, Y . Wu, J. Li, C. Yin, and X. Li, “Pseudo-label enhancement for weakly supervised object detection using self- supervised vision transformer,” Knowledge-Based Systems, vol. 311, p. 113012, 2025

    2025

  19. [27]

    Enhancing few-shot object detection through pseudo-label mining,

    P. Garcia-Fernandez, D. Cores, and M. Mucientes, “Enhancing few-shot object detection through pseudo-label mining,” Image and Vision Computing , vol. 154, p. 105379, 2025

    2025

  20. [28]

    Faster r-cnn: To- wards real-time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: To- wards real-time object detection with region proposal networks,” IEEE transactions on pattern analysis and machine intelligence, vol. 39, no. 6, pp. 1137–1149, 2016

    2016

  21. [29]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ra- manan, P. Dollár, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer vision–ECCV 2014: 13th Eu- ropean conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13 . Springe...

    2014

  22. [30]

    Yolox: Exceeding yolo series in 2021,

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Yolox: Exceeding yolo series in 2021,” arXiv preprint arXiv:2107.08430 , 2021

    2021 arXiv

  23. [31]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2980– 2988

    2017

  24. [32]

    Objects365: A large-scale, high-quality dataset for object detection,

    S. Shao, Z. Li, T. Zhang, C. Peng, G. Yu, X. Zhang, J. Li, and J. Sun, “Objects365: A large-scale, high-quality dataset for object detection,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 8430–8439

    2019

  25. [33]

    Caltech-256 object category dataset,

    G. Griffin, A. Holub, P. Perona et al. , “Caltech-256 object category dataset,” Technical Report 7694, California Institute of Technology Pasadena, Tech. Rep., 2007

    2007

  26. [34]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition . Ieee, 2009, pp. 248–255

    2009

  27. [35]

    Cifar-10 (canadian institute for advanced research),

    A. Krizhevsky, V . Nair, and G. Hinton, “Cifar-10 (canadian institute for advanced research),” URL http://www. cs. toronto. edu/kriz/cifar. html, vol. 5, no. 4, p. 1, 2010

    2010

  28. [36]

    Roboflow 100: A rich, multi-domain object detection benchmark,

    F. Ciaglia, F. S. Zuppichini, P. Guerrie, M. McQuade, and J. So- lawetz, “Roboflow 100: A rich, multi-domain object detection benchmark,” arXiv preprint arXiv:2211.13523 , 2022

    2022 arXiv

  29. [37]

    Yolov3: An incremental improve- ment,

    J. Redmon and A. Farhadi, “Yolov3: An incremental improve- ment,” arXiv preprint arXiv:1804.02767 , 2018

    2018 arXiv

  30. [38]

    Faster r-cnn: Towards real-time object detection with region proposal networks,

    S. Ren, “Faster r-cnn: Towards real-time object detection with region proposal networks,” arXiv preprint arXiv:1506.01497 , 2015

    2015 arXiv

  31. [39]

    Libra r-cnn: Towards balanced learning for object detection,

    J. Pang, K. Chen, J. Shi, H. Feng, W. Ouyang, and D. Lin, “Libra r-cnn: Towards balanced learning for object detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 821–830

    2019

Pith tools

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