Pith. sign in

REVIEW 4 major objections 5 minor 52 references

A Dynamic Transformer Network for Vehicle Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read DTNet, a vehicle detector built on YOLOv7 with dynamic weights, mixed attention, and location-aware convolution, is claimed to outperform prior methods on UA-DETRAC, BDD100K, and BIT-Vehicle.

desk verdict An honest but under-verified integration of three published modules into YOLOv7; the reported mAP margins are not trustworthy until the baselines are re-run under one protocol. read the letter →

arxiv 2506.02765 v1 pith:BOYRC64B submitted 2025-06-03 cs.CV

classification cs.CV
keywords vehicledetectiondynamicconvolutionmixedattentionTransformertranslation-variantYOLOv7UA-DETRACBDD100K
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

DTNet, a vehicle detection network built on YOLOv7, is proposed with the aim of adapting a detector to varied traffic scenes by generating weights dynamically per input. The paper argues that a dynamic convolution block, a mixed attention block combining channel attention with window-based self-attention, and a translation-variant convolution that uses spatial location information each contribute to extracting salient vehicle information. Experimental results report higher mAP than comparison methods on three datasets: 69.0% on UA-DETRAC, 57.3% on BDD100K, and 84.7% on BIT-Vehicle. The paper is trying to establish that input-adaptive mechanisms in these three forms yield a practical improvement in vehicle detection.

What carries the argument

The load-bearing mechanism is the insertion of three blocks into the YOLOv7 backbone: a dynamic convolutional layer (DCL) that generates per-input weights from spatial and channel context, a mixed attention block (MAB) that runs channel attention alongside window-based self-attention to relate channels and pixels, and a translation-variant convolution (TVConv) whose weights are conditioned on spatial location. The paper's argument is that these blocks separately handle scene adaptability, salient information extraction, and within-image structural differences.

What would settle it

Re-run all baselines, including YOLOv7, under the paper's reported training and evaluation settings on BDD100K and the other datasets, using the same evaluation code and thresholds; if YOLOv7 matches or exceeds DTNet's mAP under a unified protocol, the claimed margins collapse.

Watch

Extended reading notes

Core claim

The central claim is that DTNet is competitive for vehicle detection, stated through mAP comparisons with prior methods: 69.0% versus 67.7% for DC on UA-DETRAC, 57.3% versus 52.7% for Modified YOLOv4 on BDD100K, and 84.7% versus 81.3% for YOLOv3 on BIT-Vehicle. The comparison tables also show DTNet ahead of YOLOv7 on UA-DETRAC (69.0% versus 67.3%) and on BIT-Vehicle (84.7% versus 77.0%). Ablations on BIT-Vehicle attribute the gains to the proposed components: removing dynamic convolution, mixed attention, and translation-variant convolution drops mAP from 84.7% to 77.1%.

Load-bearing premise

The quoted mAP values for comparison methods were computed under the same protocol as DTNet (same evaluation code, class set, input resolution, confidence threshold, and data split), although the paper does not describe a unified evaluation harness and does not report YOLOv7 on BDD100K, the very backbone it modifies.

Editorial extensions

If this is right

  • If the reported margins hold under a unified protocol, DTNet offers a stronger one-stage detector for traffic surveillance and autonomous-driving settings that face occlusion and lighting variation.
  • The three-block design (dynamic weights, mixed attention, location-aware convolution) gives a template for adapting other single-stage detectors beyond vehicle detection.
  • Because DTNet builds directly on YOLOv7, practitioners can swap the backbone blocks into existing YOLOv7 deployments without redesigning the detection head.
  • The mAP gains reported on three datasets of different scale suggest the method generalizes across dataset conditions, not just one benchmark.

Reading between the lines

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

  • The BDD100K comparison omits YOLOv7, so the 4.6-point gain over Modified YOLOv4 may be partly a backbone effect; a direct YOLOv7 baseline on BDD100K is needed to isolate the proposed blocks' contribution.
  • Because the paper reports single runs and no variance, the 1.3-point margin over DC on UA-DETRAC and the 3.4-point margin over YOLOv3 on BIT-Vehicle should be treated as point estimates until repeated-seed experiments confirm them.
  • The authors note performance may drop in sunny and snowy conditions; weather-stratified evaluation on BDD100K would directly test the claimed robustness to varying environments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes DTNet, a YOLOv7-based vehicle detector that inserts three components: a dynamic convolution block (DCB) that generates input-dependent weights, a mixed information retrieval block (MIRB) containing a mixed attention block (MAB) with channel attention and window-based self-attention, and a compensation block (CB) based on translation-variant convolution (TVConv). The reported experiments on UA-DETRAC, BDD100K, and BIT-Vehicle claim consistent mAP improvements over prior detectors, e.g., 69.0% vs. 67.7% over DC on UA-DETRAC, 57.3% vs. 52.7% over Modified YOLOv4 on BDD100K, and 84.7% vs. 81.3% over YOLOv3 on BIT-Vehicle. The paper includes qualitative detection examples, precision-recall curves, and a public code link.

Significance. If the empirical comparisons hold, the contribution is a modest but useful engineering result: modular, transferable components added to a strong detector, with reported gains of 1.3--4.6 mAP points over the compared baselines and explicit attention to occlusion- and illumination-related failure cases. The paper's strengths include its release of source code, the use of three established vehicle detection datasets, and the fact that the architecture components are all grounded in independently published prior work. No circular reasoning is apparent. The main weakness is experimental: the claimed margins over baselines are not yet reliable because the evaluation protocol is underspecified, the backbone numbers are internally inconsistent, and several load-bearing comparisons are missing or incomplete.

major comments (4)
  1. [Section IV.D, Tables III and V] The paper's headline claim is the mAP advantage over prior methods, but the backbone numbers are inconsistent. Table III lists YOLOv7 at 67.3% mAP on UA-DETRAC, while Table V, also described as mAP at threshold 0.5 on the same dataset, lists YOLOv7 at 65.3%; YOLOv5s is 55.1% in Table III and 55.3% in Table V. Since the claimed gain over the strongest baseline (DC, 67.7%) is only 1.3 points, this 2.0-point discrepancy in the backbone's own score is load-bearing. Please rerun all baselines with the same evaluation code and report the numbers under a single protocol, or clearly explain why the two tables use different evaluation conditions.
  2. [Section IV.B and Tables II--IV] The comparison assumes that all quoted mAP values are protocol-compatible, but the manuscript does not describe the evaluation harness: evaluation code, class set (UA-DETRAC has four classes while BDD100K has ten; it is not stated whether mAP is computed on vehicle classes only), input resolution, confidence threshold, NMS settings, data split, or whether baseline numbers come from original papers or from re-runs. Moreover, on BDD100K the backbone YOLOv7 is absent from Table II, so the 4.6-point gain over Modified YOLOv4 cannot be separated from the improvement contributed by the backbone itself. Please re-run at least YOLOv7 and Modified YOLOv4 under the same protocol and add them to Table II.
  3. [Section IV.C, Table I] The ablation study is incomplete relative to its own text. Section IV.C states that 'DTNet without MAB' has a higher mAP than 'DTNet without MAB and TVConv', but no row named 'DTNet without MAB' appears in Table I. The table also lacks a row that removes only the dynamic convolution block, so the contribution of DCL is only inferred from the difference between 'DTNet without MAB and TVConv' (79.2%) and 'DTNet without DCL, MAB, and TVConv' (77.1%), which conflates DCL with the surrounding convolutions. In addition, with a single run and no variance measure, the reported MAB gain (0.2 points) and TVConv gain (0.2 points) are within plausible run-to-run variation. Please complete the ablation grid and report repeated-run mean plus standard deviation, or otherwise justify the stability of these differences.
  4. [Section III.A and Section IV.D] The introduction and contribution statements emphasize robustness under different lighting, occlusion, and weather conditions, but the supporting evidence is limited to three qualitative scenes and aggregate mAP values. The manuscript's own limitation statement in Section IV.D acknowledges that 'due to effects of varying environments, i.e., sunny and snowy, performance of the proposed DTNet may decrease.' Since UA-DETRAC provides per-condition annotations (cloudy, night, sunny, rainy) and the datasets include occluded vehicles, please provide per-condition mAP or per-occlusion-level results to substantiate the robustness claim, or revise the claim to focus on aggregate detection accuracy.
minor comments (5)
  1. [Section IV.D, Table V] The header 'Percision' should be 'Precision', and the definition of mAP at threshold 0.5 versus 'mAP with varying thresholds' (average over 0.50:0.05:0.95) should be stated in the metric description of Section IV.A rather than only in Section IV.D.
  2. [Section III.A, Fig. 1] The block labels in Fig. 1 are difficult to read (for example, the repeated 'MPCM' labels and the placement of 'DCL+BN+SiLU' do not clearly indicate where DCB, MIRB, CB, and DH sit in the YOLOv7 backbone); please add a clean component-to-backbone mapping, either in the figure or in a table.
  3. [References] Reference [51] is cited in the text as EfficientNet but the listed entry is PanNet (Yang et al., pansharpening); please correct either the citation or the reference list, and check that all baseline citations correspond to the models actually compared.
  4. [Section IV.B] The hardware description 'AMD EPYC of 7502P/3.35GHz and CPU of 33 cores' appears truncated, and the GPU/CUDA/cuDNN versions are not precisely described; please provide complete hardware and software specifications for reproducibility.
  5. [Throughout] The manuscript needs extensive language editing; examples include 'DTNet is superior to vehicle detection', 'can be conducted as follows', 'overcomed', 'in the filed of', and inconsistent capitalization of 'Table' and 'TABLE'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: DTNet is an empirical architecture composition evaluated against external benchmarks, with all components cited from independent prior work.

full rationale

The paper does not derive its performance from its inputs by construction. The method is presented as a composition of external, independently published components — dynamic convolution from Ref. [30], ELAN and the detection head from YOLOv7 [29], the mixed attention block from Ref. [33], and TVConv from Ref. [34] — assembled into equations (1)–(6). These equations define the forward path of the network; they are a description of an architecture, not a derivation that presupposes the claimed mAP values. The central claim is an empirical comparison on UA-DETRAC, BDD100K, and BIT-Vehicle against published baselines, which is a measurement rather than a reduction. No fitted parameter is renamed as a prediction: the ablations in Table I attribute gains to removing DCL, MAB, or TVConv, and even these gains are small (e.g., 84.5% to 84.7%) and are reported as experimental observations. There are no load-bearing self-citations: the reference list contains no cited uniqueness theorem or prior result by the present authors that is invoked to forbid alternative designs. The reviewer-identified issues — the inconsistent YOLOv7 mAP values between Tables III and V (67.3% vs 65.3%), the absence of YOLOv7 from the BDD100K table, and the unshown 'DTNet without MAB' row referenced in the text — are correctness, reproducibility, and protocol-consistency concerns, not circularity. Likewise, the paper's own limitation that performance may decrease under sunny and snowy conditions is an honest caveat and does not indicate that any conclusion is equivalent to its assumptions. Because every load-bearing component is cited from independent external work and the headline result is an external-benchmark measurement, the circularity score is 0.

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

The paper introduces no theoretical free parameters or invented entities; it is an empirical architecture study. The central claim depends on the external validity of the cited modules (DCL, MAB, TVConv), the comparability of baseline results from heterogeneous sources, and the correct use of standard dataset splits and metrics.

assumptions (3)
  • domain assumption The cited modules (DCL [30], MAB [33], TVConv [34]) work as described when inserted into YOLOv7.
    The paper uses these components without deriving them; any instability in porting them would change the measured performance.
  • domain assumption The baseline mAP values collected from different papers are directly comparable with DTNet's measurements.
    Baselines in Tables II to IV come from different sources with potentially different settings; the paper does not re-run them under a unified protocol.
  • domain assumption The standard dataset splits and evaluation tools are used correctly.
    The paper does not specify the exact evaluation code, class filtering, or threshold conventions, so correct metric computation is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Dynamic Transformer Network for Vehicle Detection." pith.science (2026). https://pith.science/paper/BOYRC64B

@misc{pith2026250602765,
  author       = {Pith},
  title        = {Pith review of: A Dynamic Transformer Network for Vehicle Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BOYRC64B}},
  note         = {Machine review of arXiv:2506.02765}
}
read the original abstract

Stable consumer electronic systems can assist traffic better. Good traffic consumer electronic systems require collaborative work between traffic algorithms and hardware. However, performance of popular traffic algorithms containing vehicle detection methods based on deep networks via learning data relation rather than learning differences in different lighting and occlusions is limited. In this paper, we present a dynamic Transformer network for vehicle detection (DTNet). DTNet utilizes a dynamic convolution to guide a deep network to dynamically generate weights to enhance adaptability of an obtained detector. Taking into relations of different information account, a mixed attention mechanism based channel attention and Transformer is exploited to strengthen relations of channels and pixels to extract more salient information for vehicle detection. To overcome the drawback of difference in an image account, a translation-variant convolution relies on spatial location information to refine obtained structural information for vehicle detection. Experimental results illustrate that our DTNet is competitive for vehicle detection. Code of the proposed DTNet can be obtained at https://github.com/hellloxiaotian/DTNet.

Figures

Figures reproduced from arXiv: 2506.02765 by the authors.

Figure 1
Figure 1. Architecture of DTNet. well as DTNet as shown in Fig.1. According to Refs. [28], we choose a modular way to express the proposed method. That is, DTNet is composed of a dynamic convolutional block (DCB), a mixed information retrieval block (MIRB), compen￾sation block (CB) and detection head (DH). DCB is used to dynamically adjust parameters to enhance the robustness of an obtained classifier for different scenes. To… view at source ↗
Figure 2
Figure 2. Visual effects of three detection methods on a scene from the UA-DETRAC: (a) YOLOv5s, (b) YOLOv7 and (c) DTNet (Ours). (a) (b) (c) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Visual effects of three detection methods on a scene from the UA-DETRAC: (a) YOLOv5s, (b) YOLOv7 and (c) DTNet (Ours). TABLE I MAP OF SERVAL METHODS ON THE BIT VEHICLE DATASET. Methods mAP DTNet without DCL, MAB and TVConv 77.1% DTNet without MAB and TVConv 79.2% DTNet without TVConv 84.5% DTNet (Ours) 84.7% TABLE II PERFORMANCE OF DIFFERENT METHODS ON THE BDD100K DATASET. Methods mAP YOLOv3 25.8% SSD 33.9% WLOD 34.… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visual effects of different detection methods on a scene from the UA-DETRAC: (a) YOLOv5s, (b) YOLOv7, and (c) DTNet (Ours). TABLE V PRECISION, RECALL, MAP AND MAP WITH VARYING THRESHOLDS OF DIFFERENT METHODS ON THE UA-DETRAC DATASET. Methods Precision Recall mAP mAP wi…
Figure 5
Figure 5. Figure 5: The PR curve of of different detection methods on the UA-DETRAC: (a) YOLOv7 and (b) DTNet (Ours). Qualitative analysis uses different methods, i.e., YOLOv5s, YOLOv7, and our DTNet on a scene to accurately detect more vehicles to test their performance in vehicle detect…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 47 canonical work pages

  1. [30]

    Adaptive dynamic filtering network for image denoising,

    H. Shen, Z.-Q. Zhao, and W. Zhang, “Adaptive dynamic filtering network for image denoising,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 2. AAAI Press, 2023, pp. 2227– 2235

  2. [33]

    Activating more pixels in image super-resolution transformer,

    X. Chen, X. Wang, J. Zhou, Y . Qiao, and C. Dong, “Activating more pixels in image super-resolution transformer,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2023, pp. 22 367–22 377

  3. [34]

    Tvconv: Efficient translation variant convolution for layout-aware visual process- ing,

    J. Chen, T. He, W. Zhuo, L. Ma, S. Ha, and S.-H. G. Chan, “Tvconv: Efficient translation variant convolution for layout-aware visual process- ing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2022, pp. 12 548–12 558

  4. [1]

    Smart traffic monitoring through real-time moving vehicle detection using deep learning via aerial images for consumer application,

    A. Singh, M. Z. U. Rahma, P. Rani, R. Sharma, E. Kaririet al., “Smart traffic monitoring through real-time moving vehicle detection using deep learning via aerial images for consumer application,”IEEE Transactions on Consumer Electronics, vol. 70, no. 4, pp. 7302–7309, 2024

  5. [2]

    Monoli: Precise monocular 3-d object detection for next-generation consumer electronics for autonomous electric vehicles,

    H. Gao, X. Yu, Y . Xu, J. Y . Kim, and Y . Wang, “Monoli: Precise monocular 3-d object detection for next-generation consumer electronics for autonomous electric vehicles,”IEEE Transactions on Consumer Electronics, vol. 70, no. 1, pp. 3475–3486, 2024

  6. [3]

    Histograms of oriented gradients for human detection,

    N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 1. IEEE, 2005, pp. 886–893

  7. [4]

    Vehicle detection, tracking and classification in urban traffic,

    Z. Chen, T. Ellis, and S. A. Velastin, “Vehicle detection, tracking and classification in urban traffic,” in2012 15th International IEEE Conference on Intelligent Transportation Systems. IEEE, 2012, pp. 951–956

  8. [5]

    Vision-based target detection and localization via a team of cooperative uav and ugvs,

    S. Minaeian, J. Liu, and Y .-J. Son, “Vision-based target detection and localization via a team of cooperative uav and ugvs,”IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 46, no. 7, pp. 1005– 1016, 2015

Show all 52 references
  1. [6]

    Detection and classification of vehicles from omnidirectional videos using multiple silhouettes,

    H. C. Karaimer, I. Baris, and Y . Bastanlar, “Detection and classification of vehicles from omnidirectional videos using multiple silhouettes,” Pattern Analysis and Applications, vol. 20, pp. 893–905, 2017

  2. [7]

    Online boosting for vehicle detection,

    W.-C. Chang and C.-W. Cho, “Online boosting for vehicle detection,” IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cyber- netics), vol. 40, no. 3, pp. 892–902, 2009

  3. [8]

    Moving vehicle detection for automatic traffic monitoring,

    J. Zhou, D. Gao, and D. Zhang, “Moving vehicle detection for automatic traffic monitoring,”IEEE Transactions on Vehicular Technology, vol. 56, no. 1, pp. 51–59, 2007

  4. [9]

    Real-time traffic sign recognition based on a general purpose gpu and deep-learning,

    K. Lim, Y . Hong, Y . Choi, and H. Byun, “Real-time traffic sign recognition based on a general purpose gpu and deep-learning,”PLoS one, vol. 12, no. 3, p. e0173317, 2017

  5. [10]

    Centernet: Keypoint triplets for object detection,

    K. Duan, S. Bai, L. Xie, H. Qi, Q. Huang, and Q. Tian, “Centernet: Keypoint triplets for object detection,” inProceedings of the IEEE/CVF International Conference On Computer Vision, 2019, pp. 6569–6578

  6. [11]

    Littleyolo-spp: A delicate real-time vehicle detection algorithm,

    E. Raniet al., “Littleyolo-spp: A delicate real-time vehicle detection algorithm,”Optik, vol. 225, p. 165818, 2021

  7. [12]

    Sinet: A scale-insensitive convolutional neural network for fast vehicle detection,

    X. Hu, X. Xu, Y . Xiao, H. Chen, S. He, J. Qin, and P.-A. Heng, “Sinet: A scale-insensitive convolutional neural network for fast vehicle detection,” IEEE Transactions on Intelligent Transportation Systems, vol. 20, no. 3, pp. 1010–1019, 2018

  8. [13]

    Vision-based vehicle detection and counting system using deep learning in highway scenes,

    H. Song, H. Liang, H. Li, Z. Dai, and X. Yun, “Vision-based vehicle detection and counting system using deep learning in highway scenes,” European Transport Research Review, vol. 11, no. 1, pp. 1–16, 2019. 8

  9. [14]

    A lightweight vehicles detection network model based on yolov5,

    X. Dong, S. Yan, and C. Duan, “A lightweight vehicles detection network model based on yolov5,”Engineering Applications of Artificial Intelligence, vol. 113, p. 104914, 2022

  10. [15]

    Multi-modality sensing and data fusion for multi-vehicle detection,

    D. Roy, Y . Li, T. Jian, P. Tian, K. R. Chowdhury, and S. Ioannidis, “Multi-modality sensing and data fusion for multi-vehicle detection,” IEEE Transactions on Multimedia, 2022

  11. [16]

    Real-time vehicle detection based on improved yolo v5,

    Y . Zhang, Z. Guo, J. Wu, Y . Tian, H. Tang, and X. Guo, “Real-time vehicle detection based on improved yolo v5,”Sustainability, vol. 14, no. 19, p. 12274, 2022

  12. [17]

    Improved faster rcnn based on feature amplification and oversampling data augmentation for oriented vehicle detection in aerial images,

    N. Mo and L. Yan, “Improved faster rcnn based on feature amplification and oversampling data augmentation for oriented vehicle detection in aerial images,”Remote Sensing, vol. 12, no. 16, p. 2558, 2020

  13. [18]

    Preprocessed faster rcnn for vehicle detection,

    M. Manana, C. Tu, and P. A. Owolawi, “Preprocessed faster rcnn for vehicle detection,” in2018 International Conference on Intelligent and Innovative Computing Applications. IEEE, 2018, pp. 1–4

  14. [19]

    Multeye: Monitoring system for real-time vehicle detection, tracking and speed estimation from uav imagery on edge-computing platforms,

    N. Balamuralidhar, S. Tilon, and F. Nex, “Multeye: Monitoring system for real-time vehicle detection, tracking and speed estimation from uav imagery on edge-computing platforms,”Remote sensing, vol. 13, no. 4, p. 573, 2021

  15. [20]

    Performance enhancement of yolov3 by adding prediction layers with spatial pyramid pooling for vehicle detection,

    K.-J. Kim, P.-K. Kim, Y .-S. Chung, and D.-H. Choi, “Performance enhancement of yolov3 by adding prediction layers with spatial pyramid pooling for vehicle detection,” in2018 15th IEEE International Confer- ence on Advanced Video and Signal Based Surveillance. IEEE, 2018, pp. 1–6

  16. [21]

    Deep neural network based vehicle detection and classification of aerial images

    S. Kumar, A. Jain, S. Rani, H. Alshazly, S. A. Idris, and S. Bourouis, “Deep neural network based vehicle detection and classification of aerial images.”Intelligent Automation & Soft Computing, vol. 34, no. 1, 2022

  17. [22]

    Traffic management: Multi-scale vehicle detection in varying weather condi- tions using yolov4 and spatial pyramid pooling network,

    M. Humayun, F. Ashfaq, N. Z. Jhanjhi, and M. K. Alsadun, “Traffic management: Multi-scale vehicle detection in varying weather condi- tions using yolov4 and spatial pyramid pooling network,”Electronics, vol. 11, no. 17, p. 2748, 2022

  18. [23]

    A resource- efficient cnn-based method for moving vehicle detection,

    Z. Charouh, A. Ezzouhri, M. Ghogho, and Z. Guennoun, “A resource- efficient cnn-based method for moving vehicle detection,”Sensors, vol. 22, no. 3, p. 1193, 2022

  19. [24]

    Dynamic convolution: Attention over convolution kernels,

    Y . Chen, X. Dai, M. Liu, D. Chen, L. Yuan, and Z. Liu, “Dynamic convolution: Attention over convolution kernels,” inProceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. IEEE, 2020, pp. 11 030–11 039

  20. [25]

    Dynamic convolutions: Exploiting spatial sparsity for faster inference,

    T. Verelst and T. Tuytelaars, “Dynamic convolutions: Exploiting spatial sparsity for faster inference,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition. IEEE, 2020, pp. 2320–2329

  21. [26]

    Dynamic region-aware convolution,

    J. Chen, X. Wang, Z. Guo, X. Zhang, and J. Sun, “Dynamic region-aware convolution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2021, pp. 8064–8073

  22. [27]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2018, pp. 7132–7141

  23. [28]

    Residual dense network for image restoration,

    Y . Zhang, Y . Tian, Y . Kong, B. Zhong, and Y . Fu, “Residual dense network for image restoration,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 7, pp. 2480–2495, 2020

  24. [29]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2023, pp. 7464–7475

  25. [31]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inInternational Conference on Machine Learning. pmlr, 2015, pp. 448–456

  26. [32]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,

    S. Elfwing, E. Uchibe, and K. Doya, “Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,” Neural Networks, vol. 107, pp. 3–11, 2018

  27. [35]

    Ua-detrac: A new benchmark and protocol for multi-object detection and tracking,

    L. Wen, D. Du, Z. Cai, Z. Lei, M.-C. Chang, H. Qi, J. Lim, M.-H. Yang, and S. Lyu, “Ua-detrac: A new benchmark and protocol for multi-object detection and tracking,”Computer Vision and Image Understanding, vol. 193, p. 102907, 2020

  28. [36]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2020, pp. 2636–2645

  29. [37]

    Vehicle type classification using a semisupervised convolutional neural network,

    Z. Dong, Y . Wu, M. Pei, and Y . Jia, “Vehicle type classification using a semisupervised convolutional neural network,”IEEE Transactions on Intelligent Transportation Systems, vol. 16, no. 4, pp. 2247–2256, 2015

  30. [38]

    Stochastic gradient descent tricks,

    L. Bottou, “Stochastic gradient descent tricks,” inNeural Networks: Tricks of the Trade: Second Edition. Springer, 2012, pp. 421–436

  31. [39]

    Stevens, L

    E. Stevens, L. Antiga, and T. Viehmann,Deep Learning with PyTorch: Build, train, and tune neural networks using Python tools. Manning, 2020

  32. [40]

    Yolov3: An incremental improvement,

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

  33. [41]

    Ssd: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “Ssd: Single shot multibox detector,” inin Proceedings of European Conference on Computer Vision. Springer, 2016, pp. 21–37

  34. [42]

    Wasserstein loss-based deep object detection,

    Y . Han, X. Liu, Z. Sheng, Y . Ren, X. Han, J. You, R. Liu, and Z. Luo, “Wasserstein loss-based deep object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. IEEE, 2020, pp. 998–999

  35. [43]

    Yolov4: Op- timal speed and accuracy of object detection,

    A. Bochkovskiy, C.-Y . Wang, and H.-Y . M. Liao, “Yolov4: Op- timal speed and accuracy of object detection,”arXiv preprint arXiv:2004.10934, 2020

  36. [44]

    A deep learning-based hybrid framework for object detection and recognition in autonomous driving,

    Y . Li, H. Wang, L. M. Dang, T. N. Nguyen, D. Han, A. Lee, I. Jang, and H. Moon, “A deep learning-based hybrid framework for object detection and recognition in autonomous driving,”IEEE Access, vol. 8, pp. 194 228–194 239, 2020

  37. [45]

    Improved vision-based vehicle detection and classification by optimized yolov4,

    J. Zhao, S. Hao, C. Dai, H. Zhang, L. Zhao, Z. Ji, and I. Ganchev, “Improved vision-based vehicle detection and classification by optimized yolov4,”IEEE Access, vol. 10, pp. 8590–8603, 2022

  38. [46]

    ultralytics/yolov5: v3. 0,

    G. Jocher, A. Stoken, J. Borovec, L. Changyu, A. Hogan, L. Diaconu, J. Poznanski, L. Yu, P. Rai, R. Ferridayet al., “ultralytics/yolov5: v3. 0,”Zenodo, 2020

  39. [47]

    Small object detection in traffic scenes based on attention feature fusion,

    J. Lian, Y . Yin, L. Li, Z. Wang, and Y . Zhou, “Small object detection in traffic scenes based on attention feature fusion,”Sensors, vol. 21, no. 9, p. 3031, 2021

  40. [48]

    Vehicle target detection algorithm based on yolov5,

    X. He, “Vehicle target detection algorithm based on yolov5,”Frontiers in Computing and Intelligent Systems, vol. 3, no. 1, pp. 56–59, 2023

  41. [49]

    Densenet: Implementing efficient convnet descriptor pyra- mids,

    F. Iandola, M. Moskewicz, S. Karayev, R. Girshick, T. Darrell, and K. Keutzer, “Densenet: Implementing efficient convnet descriptor pyra- mids,”arXiv preprint arXiv:1404.1869, 2014

  42. [50]

    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,” inProceedings of the IEEE International Conference on Computer Vision, 2017, pp. 2980–2988

  43. [51]

    Pannet: A deep network architecture for pan-sharpening,

    J. Yang, X. Fu, Y . Hu, Y . Huang, X. Ding, and J. Paisley, “Pannet: A deep network architecture for pan-sharpening,” inProceedings of the IEEE International Conference on Computer Vision. IEEE, 2017, pp. 5449–5457

  44. [52]

    A cross-scale and illumination invariance-based model for robust object detection in traffic surveillance scenarios,

    Y .-F. Lu, J.-W. Gao, Q. Yu, Y . Li, Y .-S. Lv, and H. Qiao, “A cross-scale and illumination invariance-based model for robust object detection in traffic surveillance scenarios,”IEEE Transactions on Intelligent Trans- portation Systems, 2023

Pith tools

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