Pith. sign in

REVIEW 5 major objections 8 minor 30 references

OD-VIRAT: A Large-Scale Benchmark for Object Detection in Realistic Surveillance Environments

T0 review · 5 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper presents two surveillance object detection benchmarks, OD-VIRAT Large and OD-VIRAT Tiny, built from VIRAT footage with 8.7 million and 288,901 annotated instances respectively, and reports Deformable-DETR with ResNet50 as the…

desk verdict A useful repackaged surveillance benchmark whose headline leaderboard claim is internally contradicted by its own table. read the letter →

arxiv 2507.12396 v2 pith:EHBUG7YH submitted 2025-07-16 cs.CV

Hayat Ullah , Abbas Khan , Arslan Munir , Hari Kalva This is my paper
classification cs.CV
keywords objectdetectionbenchmarkdatasetsurveillancevideoVIRATsmallDeformable-DETRCOCOformatcorruptionrobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a large-scale, realistic surveillance object detection benchmark can be built by converting VIRAT Ground 2.0 event-recognition videos into frame-level detection datasets, and that this resource supports meaningful comparison of modern detectors under challenging elevated-view, small-object conditions. If the datasets are sound, researchers gain a much-needed surveillance-specific benchmark with 8.7 million labeled instances in about 600,000 frames (Large) and 288,901 instances in 19,860 frames (Tiny). The accompanying experiments on Tiny report Deformable-DETR with a ResNet50 backbone as the strongest detector by mean average precision, with RetinaNet close behind at lower cost, and document how all five architectures degrade under common corruptions.

What carries the argument

The carrying device is the annotation-conversion pipeline: bounding-box coordinates are retrieved frame-by-frame from VIRAT's objects.txt, compiled by matching frame IDs and overlaying boxes onto frames, converted from .txt to COCO .json, and verified by visual application to the images. Frame-skip frequency is the dial that produces the two datasets: 0 frame-skip yields the 599,996-frame OD-VIRAT Large, and 30 frame-skip yields the 19,860-frame OD-VIRAT Tiny. This pipeline makes the benchmark's validation claims stand or fall on the quality and alignment of the inherited VIRAT annotations.

What would settle it

Sample, say, 200 frames from OD-VIRAT Large and have fresh annotators draw boxes for the five classes; measure IoU between the retrieved boxes and the re-drawn ones, and also compare the scene composition of the 69 dropped VIRAT videos against the 10 retained scenes. Low agreement on small objects, or systematically different scenes in the dropped set, would show the benchmark inherits annotation errors or a biased scene sample.

Watch

Extended reading notes

Core claim

The central claim is that OD-VIRAT Large and OD-VIRAT Tiny are ready-to-use object detection benchmarks derived from VIRAT's existing surveillance footage and annotations, spanning ten scenes and five classes (bike/bicycle, car, carrying object, person, vehicle) with frames converted from videos using 0 and 30 frame-skip strategies. On the Tiny test set, the paper reports that Deformable-DETR with ResNet50 achieves the best mAP (up to 75.0 with two-stage and iterative bounding box refinement, trained with pretrained weights), that RetinaNet with ResNet backbones is a strong second, and that YOLOX leads in frames per second. The paper presents these results as baseline references for future surveillance detection research.

Load-bearing premise

The entire benchmark inherits the original VIRAT bounding boxes and class labels without re-annotation, so the claimed 8.7 million accurate instances depends on objects.txt being complete, correctly aligned to frames, and correctly labeled.

Editorial extensions

If this is right

  • OD-VIRAT Large provides 8.7 million instances across 599,996 images, exceeding the image count of commonly used general detection benchmarks and targeting elevated-view, small-object surveillance conditions.
  • Deformable-DETR with ResNet50 (two-stage, with iterative refinement, pretrained) sets the top baseline on OD-VIRAT Tiny test at 75.0 mAP, while RetinaNet with ResNet18 offers competitive mAP75 and mAPM at a fraction of the parameters and GFLOPs.
  • YOLOX achieves the highest throughput at 256.6 FPS on average, making it the practical real-time choice among the tested detectors.
  • All evaluated detectors lose substantial accuracy under Gaussian noise, motion blur, snow, and elastic transform, with DETR dropping below 10% mAP at severity 5 on some corruptions, indicating that robustness to surveillance-specific image degradation remains an open problem.

Reading between the lines

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

  • Because the boxes come from VIRAT's event-oriented objects.txt without re-annotation, the benchmark's accuracy ceiling is set by that file: a manual re-annotation spot-check on sampled frames would be the natural validation step and would give the community a quality estimate.
  • The 30-frame-skip Tiny set likely overrepresents persistent objects (those visible across many frames) compared with fleeting ones, so class-frequency and object-duration statistics should be checked before using Tiny to study rare or transient object classes.
  • The same conversion pipeline could be applied to other surveillance corpora with bounding-box annotation files, and the retained frame IDs would allow temporal tasks such as re-identification or tracking to be layered onto the benchmark.
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 / 8 minor

Summary. The paper introduces two object detection benchmarks, OD-VIRAT Large and OD-VIRAT Tiny, derived from the VIRAT Ground 2.0 surveillance dataset. The authors convert VIRAT's event-oriented bounding-box annotations into COCO-format detection annotations, select 260 of the original 329 videos covering 10 scenes, and create Large via 0-frame-skip conversion (599,996 images, 8.7 million instances) and Tiny via 30-frame-skip conversion (19,860 images, 288,901 instances). The paper then benchmarks RTMDET, YOLOX, RetinaNet, DETR, and Deformable-DETR on the Tiny test set, reporting mAP, robustness under image corruptions, batch-size sensitivity, and FPS. The headline experimental claim is that Deformable-DETR with a ResNet50 backbone achieves the best mAP on OD-VIRAT Tiny.

Significance. If the dataset is sound, it is a potentially valuable resource: it repurposes a widely used surveillance video corpus into a large-scale detection benchmark, provides public code and models, and offers an extensive empirical comparison on surveillance imagery. The scale of OD-VIRAT Large, if validated, would be a meaningful addition to the small set of realistic surveillance detection datasets. The paper also ships a reproducible conversion pipeline, which is a concrete strength. However, the significance is currently limited by two factors: the flagship Large benchmark is never evaluated, and the central leaderboard claim in the abstract, Fig. 1, and Section IV.C is internally contradicted by Table III. These issues must be resolved before the paper's conclusions can be accepted as stated.

major comments (5)
  1. [Section IV.C / Table III] The repeated claim that Deformable-DETR with ResNet50 obtains the best mAP values across each IoU and scale threshold is contradicted by the paper's own Table III. For example, pre-trained RetinaNet-ResNet101 achieves mAP75 = 71.6 and mAPM = 88.3, whereas Deformable-DETR-ResNet50 (pre-trained, without iterative refinement or two-stage decoding) achieves mAP75 = 71.0 and mAPM = 83.1; RetinaNet-ResNet18 also has higher mAP75 and mAPM. The claim is only true for the specific two-stage Deformable-DETR variant and only for overall mAP, not for every sub-metric. Please correct the abstract, Fig. 1, and Section IV.C to name the exact variant and to state that it wins on overall mAP, not on all IoU and scale thresholds.
  2. [Section IV.A] All object detection experiments are conducted only on OD-VIRAT Tiny; the OD-VIRAT Large benchmark is never evaluated despite being presented as the primary large-scale contribution in the abstract and introduction. This means the paper does not demonstrate that the large-scale benchmark is usable for training or that its annotation quality supports the claimed scale. Either include at least a representative training/evaluation result on OD-VIRAT Large, or explicitly reframe the paper's contribution as the Tiny benchmark plus an annotation-only release of Large, with a validation study showing that the Large annotations are complete and correct.
  3. [Section II] The exclusion of 69 of the 329 VIRAT videos is described only as due to 'incorrect temporal relation between objects and their annotations,' with no further detail. This is a load-bearing data-selection decision for a benchmark paper. Please specify the criterion, report the number of excluded videos per original scene, state which of the original 11 scenes is absent from the final 10-scene set, and show that the excluded videos do not systematically differ in scene content, object classes, or annotation density. The test set containing only 8 scenes further compounds this concern, since the reported per-scene test distribution in Fig. 4 omits two scenes.
  4. [Section II.C] The annotation pipeline retrieves bounding boxes from VIRAT's objects.txt and verifies them only by overlaying boxes onto frames; there is no re-annotation or independent quality check. Because VIRAT's annotations are event-oriented and may omit stationary objects or contain frame-alignment errors, the benchmark inherits any such incompleteness. Please provide a quantitative validation protocol: a manually inspected random sample with reported agreement, per-class bounding-box counts, the fraction of frames with zero annotations, and a discussion of whether static objects (e.g., parked cars) are included.
  5. [Table III / Section IV.C] All quantitative results are from single runs, and some leaderboard differences are small (e.g., 70.8 mAP for RetinaNet-ResNet101 vs. 71.6 mAP for Deformable-DETR-ResNet50 without refinement). Without multiple seeds or error bars, these differences are not statistically meaningful. The batch-size variation in Tables IX-XIII already shows mAP fluctuations of several points, so reporting a single run for each configuration is insufficient for the paper's ranking claims. Please run at least three seeds for the main configurations and report mean and standard deviation.
minor comments (8)
  1. [Abstract / Section I] The phrase 'first work to examine the performance of these recently published state-of-the-art object detection architectures on realistic surveillance imagery' is an overclaim; many prior studies benchmark detectors on surveillance datasets such as VisDrone. Please temper the novelty claim.
  2. [Table I] The table header contains the typo 'Resoultion' and the scene/video/image column layout is confusing because the same label row is used for three groups. Please restructure the table so that the numbers of scenes, videos, and images are separately and unambiguously labeled.
  3. [Section II] The text says the retrieved 260 videos contain visuals of 10 distinct scenes, but the original VIRAT Ground 2.0 dataset has 11 scenes. Please state which scene is dropped and why.
  4. [Section IV.C / Fig. 1] Fig. 1's caption states that Deformable-DETR with ResNet50 outperforms other counterparts, but the figure does not distinguish the two-stage variant from the plain Deformable-DETR. Please mark the exact variant in the figure.
  5. [Tables IV-VIII] The table titles use 'mP Cvalues' and 'mP C' instead of 'mPC' or 'Mean Performance under Corruption.' Please standardize the terminology and define mPC in Section IV.D.
  6. [Section IV.C] The sentence 'two-stage Deformable-DETR offers an average improvement of 3.1% and 3.8% over RetinaNet with ResNet50 and ResNet101' is ambiguous: it should specify that the comparison is against pre-trained RetinaNet variants and should report the exact metric being averaged.
  7. [Section II.C] No per-class instance counts or per-class AP results are reported. Since the benchmark has five classes, please include a class-wise breakdown in the dataset statistics and in the evaluation tables.
  8. [Section I / Conclusion] There are several typos throughout, including 'Arlsan Munir' in the author affiliation, 'varients', 'bechmarking', and 'repsectively.' A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dataset is an independent artifact converted from VIRAT, and the model rankings are measured, not derived from the paper's own claims.

full rationale

The paper's central contribution is the construction of OD-VIRAT Large and OD-VIRAT Tiny from the VIRAT Ground 2.0 dataset. Section II.C describes a data-processing pipeline — retrieving bounding-box coordinates from VIRAT's objects.txt, applying them to frames, and converting to COCO JSON — rather than a fitted model or a derivation that presupposes the paper's conclusions. No parameter is fit to a subset of data and then renamed as a prediction, and no claimed result is defined in terms of the benchmark's own output. The benchmarking experiments in Sections IV.C–IV.F evaluate externally published detectors on the new test set; reporting the best measured mAP is the intended use of a benchmark, not a circular inference. The only self-referential concern would be if the paper cited its own prior work as load-bearing support, but the relevant citations, especially VIRAT [18], are external and independent. The internal inconsistency between the Section IV.C statement that Deformable-DETR 'obtains the best mAP values across each IoU and scale threshold' and Table III, where RetinaNet-ResNet101 (pretrained) has higher mAP75 and mAPM, is a correctness and consistency issue for the experimental narrative, not a circularity issue: the claim is contradicted by, rather than reduced to, the paper's own measurements. Similarly, the exclusion of 69 videos due to 'incorrect temporal relation' raises a data-quality concern, but it does not make the benchmark's construction circular. The paper is therefore self-contained with respect to circularity, and the appropriate finding is no significant circularity (score 0).

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

No fitted parameters or invented entities. The benchmark is a reformatting of an existing dataset, so the load-bearing inputs are the quality of the original annotations and the representativeness of the selected videos.

assumptions (3)
  • domain assumption The bounding boxes in VIRAT objects.txt are complete and correctly aligned with their frames.
    Section II.C retrieves coordinates from objects.txt without re-annotation; if this fails, all benchmark statistics and model scores inherit the error.
  • domain assumption Surveillance recorded from rooftop cameras in 10 scenes is representative of realistic surveillance environments.
    Introduction and Section II assert realism; the 10 scenes, all daytime, limit generalization to night or indoor surveillance.
  • domain assumption Standard training recipes for the five detectors transfer to this dataset without tuning.
    Section IV.A uses fixed 50 epochs and default MMDetection configs; no hyperparameter search, so reported mAP may understate model potential.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OD-VIRAT: A Large-Scale Benchmark for Object Detection in Realistic Surveillance Environments." pith.science (2026). https://pith.science/paper/EHBUG7YH

@misc{pith2026250712396,
  author       = {Pith},
  title        = {Pith review of: OD-VIRAT: A Large-Scale Benchmark for Object Detection in Realistic Surveillance Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHBUG7YH}},
  note         = {Machine review of arXiv:2507.12396}
}
read the original abstract

Realistic human surveillance datasets are crucial for training and evaluating computer vision models under real-world conditions, facilitating the development of robust algorithms for human and human-interacting object detection in complex environments. These datasets need to offer diverse and challenging data to enable a comprehensive assessment of model performance and the creation of more reliable surveillance systems for public safety. To this end, we present two visual object detection benchmarks named OD-VIRAT Large and OD-VIRAT Tiny, aiming at advancing visual understanding tasks in surveillance imagery. The video sequences in both benchmarks cover 10 different scenes of human surveillance recorded from significant height and distance. The proposed benchmarks offer rich annotations of bounding boxes and categories, where OD-VIRAT Large has 8.7 million annotated instances in 599,996 images and OD-VIRAT Tiny has 288,901 annotated instances in 19,860 images. This work also focuses on benchmarking state-of-the-art object detection architectures, including RETMDET, YOLOX, RetinaNet, DETR, and Deformable-DETR on this object detection-specific variant of VIRAT dataset. To the best of our knowledge, it is the first work to examine the performance of these recently published state-of-the-art object detection architectures on realistic surveillance imagery under challenging conditions such as complex backgrounds, occluded objects, and small-scale objects. The proposed benchmarking and experimental settings will help in providing insights concerning the performance of selected object detection models and set the base for developing more efficient and robust object detection architectures.

Figures

Figures reproduced from arXiv: 2507.12396 by the authors.

Figure 1
Figure 1. Model Complexity vs Accuracy (mAP) trade-off com￾parison: We evaluate the performance of five main-stream object detection architectures on OD-VIRAT Tiny dataset and compared the obtained mAP values against model complex￾ities (# of parameters). The Deformable-DETR architecture with resnet50 backbone outperform other counterparts by obtaining the best mAP value. I. INTRODUCTION O BJECT detection in surveillance envi… view at source ↗
Figure 2
Figure 2. Graphical illustration of train, validation, and test split [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Data distribution across 10 scenes in train, validation, [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Visual overview of different scenes from OD-VIRAT [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: The number of bounding boxes per frame across 10 different scenes in train, validation, and test sets of OD-VIRAT [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Convergence curves (classification loss and bbox prediction loss) curves of RTMDET, YOLOX, RetinaNet, DETR, and [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Visual comparative analysis of selected object detection models on five test images. (a) RTMDET, (b) YOLOX, (c) [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Visual comparative analysis of selected object detection models on test image perturbed with Motion Blur and Snow [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: The obtained quantitative results in terms of [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Performance reduction of each model on test images perturbed with Gaussian Noise, Motion Blur, Snow, and [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 18 canonical work pages

  1. [1]

    Anchor- free oriented proposal generator for object detection,

    G. Cheng, J. Wang, K. Li, X. Xie, C. Lang, Y . Yao, and J. Han, “Anchor- free oriented proposal generator for object detection,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–11, 2022

    work page 2022

  2. [2]

    Diffusiondet: Diffusion model for object detection,

    S. Chen, P. Sun, Y . Song, and P. Luo, “Diffusiondet: Diffusion model for object detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 830–19 843

    work page 2023

  3. [3]

    Rank-detr for high quality object detection,

    Y . Pu, W. Liang, Y . Hao, Y . Yuan, Y . Yang, C. Zhang, H. Hu, and G. Huang, “Rank-detr for high quality object detection,” Advances in Neural Information Processing Systems, vol. 36, 2024

    work page 2024

  4. [4]

    A fast accurate fine-grain object detection model based on yolov4 deep neural network,

    A. M. Roy, R. Bose, and J. Bhaduri, “A fast accurate fine-grain object detection model based on yolov4 deep neural network,” Neural Computing and Applications, pp. 1–27, 2022

    work page 2022

  5. [5]

    A dual weighting label assign- ment scheme for object detection,

    S. Li, C. He, R. Li, and L. Zhang, “A dual weighting label assign- ment scheme for object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 9387–9396

    work page 2022

  6. [6]

    Yolov6: A single-stage object detection framework for industrial applications,

    C. Li, L. Li, H. Jiang, K. Weng, Y . Geng, L. Li, Z. Ke, Q. Li, M. Cheng, W. Nie et al., “Yolov6: A single-stage object detection framework for industrial applications,” arXiv preprint arXiv:2209.02976, 2022. 14

    arXiv 2022

  7. [7]

    Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7464–7475

    work page 2023

  8. [8]

    Exploring plain vision transformer backbones for object detection,

    Y . Li, H. Mao, R. Girshick, and K. He, “Exploring plain vision transformer backbones for object detection,” in European Conference on Computer Vision. Springer, 2022, pp. 280–296

    2022

Show all 30 references
  1. [9]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer, ...

    2014

  2. [10]

    The pascal visual object classes (voc) challenge,

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

    2010

  3. [11]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning,

    F. Yu, H. Chen, X. Wang, W. Xian, Y . Chen, F. Liu, V . Madhavan, and T. Darrell, “Bdd100k: A diverse driving dataset for heterogeneous multitask learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2636–2645

    2020

  4. [12]

    Visual object classes chal- lenge 2012,

    Everingham, Mark and Van Gool, Luc and Williams, Christopher KI and Winn, John and Zisserman, Andrew, “Visual object classes chal- lenge 2012,” http://host.robots.ox.ac.uk/pascal/VOC/voc2012/, accessed on Date (May 30, 2024)

    2012

  5. [13]

    Coco: Microsoft common objects in context,

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll ´ar, Piotr and Zitnick, C Lawrence, “Coco: Microsoft common objects in context,” https://cocodataset.org/, accessed on Date (May 30, 2024)

    2024

  6. [14]

    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

  7. [15]

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale,

    A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont- Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, T. Duerig, and V . Ferrari, “The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale,” IJCV, 2020

    2020

  8. [16]

    Detection and tracking meet drones challenge,

    P. Zhu, L. Wen, D. Du, X. Bian, H. Fan, Q. Hu, and H. Ling, “Detection and tracking meet drones challenge,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–1, 2021

    2021

  9. [17]

    V3det: Vast vocabulary visual detection dataset,

    J. Wang, P. Zhang, T. Chu, Y . Cao, Y . Zhou, T. Wu, B. Wang, C. He, and D. Lin, “V3det: Vast vocabulary visual detection dataset,” in The IEEE International Conference on Computer Vision (ICCV), October 2023

    2023

  10. [18]

    A large-scale benchmark dataset for event recognition in surveillance video,

    S. Oh, A. Hoogs, A. Perera, N. Cuntoor, C.-C. Chen, J. T. Lee, S. Mukherjee, J. Aggarwal, H. Lee, L. Davis et al., “A large-scale benchmark dataset for event recognition in surveillance video,” in CVPR

  11. [19]

    Rtmdet: An empirical study of designing real-time object detectors,

    C. Lyu, W. Zhang, H. Huang, Y . Zhou, Y . Wang, Y . Liu, S. Zhang, and K. Chen, “Rtmdet: An empirical study of designing real-time object detectors,” arXiv preprint arXiv:2212.07784, 2022

    2022 arXiv

  12. [20]

    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

  13. [21]

    Focal loss for dense object detection,

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

    2017

  14. [22]

    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 European conference on computer vision. Springer, 2020, pp. 213– 229

    2020

  15. [23]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,” arXiv preprint arXiv:2010.04159, 2020

    2010 arXiv

  16. [24]

    Cspnext: A new efficient token hybrid backbone,

    X. Chen, C. Yang, J. Mo, Y . Sun, H. Karmouni, Y . Jiang, and Z. Zheng, “Cspnext: A new efficient token hybrid backbone,” Engineering Applications of Artificial Intelligence, vol. 132, p. 107886, 2024

    2024

  17. [25]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2117–2125

    2017

  18. [26]

    Cspnet: A new backbone that can enhance learning capability of cnn,

    C.-Y . Wang, H.-Y . M. Liao, Y .-H. Wu, P.-Y . Chen, J.-W. Hsieh, and I.-H. Yeh, “Cspnet: A new backbone that can enhance learning capability of cnn,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 390–391

    2020

  19. [27]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  20. [28]

    Pyramid vision transformer: A versatile backbone for dense prediction without convolutions,

    W. Wang, E. Xie, X. Li, D.-P. Fan, K. Song, D. Liang, T. Lu, P. Luo, and L. Shao, “Pyramid vision transformer: A versatile backbone for dense prediction without convolutions,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 568–578

    2021

  21. [29]

    MMDetection: Open mmlab detection toolbox and benchmark,

    K. Chen, J. Wang, J. Pang, Y . Cao, Y . Xiong, X. Li, S. Sun, W. Feng, Z. Liu, J. Xu, Z. Zhang, D. Cheng, C. Zhu, T. Cheng, Q. Zhao, B. Li, X. Lu, R. Zhu, Y . Wu, J. Dai, J. Wang, J. Shi, W. Ouyang, C. C. Loy, and D. Lin, “MMDetection: Open mmlab detection toolbox and benchmar...

    1906 arXiv

  22. [2011]

    3153–3160

    IEEE, 2011, pp. 3153–3160

    2011

Pith tools

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