Pith. sign in

REVIEW 3 major objections 6 minor 40 references

3A-YOLO: New Real-Time Object Detectors with Triple Discriminative Awareness and Coordinated Representations

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a YOLOv4 detector can gain 6.2% AP while keeping real-time speed by replacing its detection head with the triple-attention TDA-YOLO Module and adding coordinated channel-position attention.

desk verdict Incremental YOLOv4 attention integration with plausible but unverified gains, tripped up by a scale-awareness branch that as written just stacks the same feature map with itself. read the letter →

arxiv 2412.07168 v1 pith:K4FU5RS6 submitted 2024-12-10 cs.CV

Xuecheng Wu , Junxiao Xue , Liangyu Fu , Jiayu Nie , Danlei Huang , Xinyi Yin This is my paper · ORCID
classification cs.CV
keywords real-timeobjectdetectionYOLOattentionmechanismstriplediscriminativeawarenessCoordinateheaddesignCOCObenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that YOLO detectors leave accuracy on the table by adding attention mechanisms piecemeal rather than coordinating them in the detection head itself. Its central claim is that a redesigned head, the TDA-YOLO Module, which applies scale-aware, spatial-aware, and task-aware attention in a hierarchical sequence, together with Coordinate Attention in the intermediate layers and neck-and-training adjustments, lifts YOLOv4 from 43.5% to 49.7% AP on COCO test-dev at $608\times608$ resolution while retaining 60.1 FPS. In the ablation, the new head alone contributes the largest single gain (+1.8% AP). The paper reports that the same recipe on a larger backbone reaches 50.6% AP and that tiny and nano versions beat their direct YOLO counterparts, so the claimed effect is a general real-time detector upgrade rather than a single-model artifact.

What carries the argument

The carrying object is the TDA-YOLO Module. It stacks the first neck feature with itself to expose a level dimension $SL$, flattens spatial positions into $S$, then applies three attention components: scale-aware attention ($\xi_{SL}$) computed by global pooling, a $1\times1$ convolution, and hard sigmoid; spatial-aware attention ($\xi_S$) using deformable-convolution sparse sampling with learned offsets and importance scalars; and task-aware attention ($\xi_C$) built from an adapted DY-ReLU-A hyper function. One adapted Dynamic Block computes $W(F)=\xi_C(\xi_S(\xi_{SL}(F)\cdot F)\cdot F)\cdot F$; the TDA module stacks two such blocks, reshapes back to $C\times H\times W$, and hands to the original YOLO head. The supporting mechanism is Coordinate Attention, which pools along horizontal and vertical directions separately and multiplies the input by the two resulting weight maps, giving the head both channel relationships and precise positional information.

What would settle it

Rerun the YOLOv4 baseline and the baseline-plus-TDA-YOLO-Module configuration from Table IV with several random seeds and identical hyperparameters, then compare the mean AP gap; if the average gain is much smaller than 1.8 points or the confidence intervals overlap, the headline attribution fails. A complementary check is to add the TDA module alone to an untouched YOLOv4 model and measure its COCO test-dev AP at $608\times608$ without any other modifications.

Watch

Extended reading notes

Core claim

The discovery, on the paper's own terms, is that a standard YOLOv4 detector can be improved by 6.2% AP without losing its real-time character by replacing the plain head with the TDA-YOLO Module and coordinating the intermediate representations. The head takes features $F_i \in \mathbb{R}^{C \times H \times W}$, creates a scale-level dimension, and iterates two adapted Dynamic Blocks whose attention operators act on scale, spatial, and channel axes via $W(F) = \xi_C(\xi_S(\xi_{SL}(F)\cdot F)\cdot F)\cdot F$. Coordinate Attention is inserted at $C_3$, $C_4$, and $C_5$ to encode direction-aware positional information, and neck changes plus Mosaic augmentation, Focal Loss, label smoothing, and DIoU-NMS add further gains. The bottom-line reported numbers are $49.7\%$ AP at $608\times608$ with 60.1 FPS, $50.6\%$ AP for the scaled X version at $640\times640$, and $93.1\%$ AP50 on Pascal VOC, with the TDA module identified as the largest ablation contributor.

Load-bearing premise

The load-bearing premise is that each reported ablation increment, especially the +1.8 AP assigned to the TDA-YOLO Module, is causal and reproducible rather than an artifact of one favorable training run, one hyperparameter setting, or interactions among the stacked changes.

Editorial extensions

If this is right

  • At $608\times608$ input, 3A-YOLO reaches $49.7\%$ AP on COCO test-dev at 60.1 FPS, a 6.2-point gain over YOLOv4's $43.5\%$ AP.
  • The scaled 3A-YOLO-X reaches $50.6\%$ AP at $640\times640$ and 57.4 FPS, 3.1 points above YOLOv4-CSP.
  • In the step-by-step ablation, the TDA-YOLO Module adds 1.8 points AP by itself, and the three attention components in the module outperform every subset of them.
  • 3A-YOLO-Tiny reaches 25.9% AP on COCO val, 6.1 points above YOLOv4-Tiny, while 3A-YOLO-Nano reaches 23.5% AP with 3.42M parameters.
  • On Pascal VOC test, 3A-YOLO reaches $93.1\%$ AP50, surpassing YOLOv5-L by 3.7 points despite using a smaller backbone.

Reading between the lines

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

  • The attribution of the largest gain to the TDA-YOLO Module rests on a single training run, so the +1.8-point increment should be read as provisional until multi-seed ablations are reported.
  • The paper leaves integration on newer baselines such as YOLOv7 explicitly untested; applying the module there is the most direct check of whether the improvement transfers.
  • The module adds only about 0.8M parameters in the main ablation, which suggests the same head-based attention recipe could be transplanted to other one-stage detectors, not only the YOLO lineage.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes 3A-YOLO, a family of real-time object detectors built on YOLOv4 and YOLOv4-Tiny, whose central novelty is a TDA-YOLO Module that applies scale-, spatial-, and task-aware attention in a hierarchical manner, plus Coordinate Attention on intermediate features, neck modifications, and training tricks. The authors report substantial gains over YOLOv4 on COCO test-dev (49.7 AP at 608×608 vs. 43.5 AP, and 50.6 AP for 3A-YOLO-X at 640×640), competitive FPS, and ablation studies attributing +1.8 AP to the TDA-YOLO Module and +1.0 AP50 to the scale-aware component alone. The paper also reports results on Pascal VOC and for Tiny/Nano variants.

Significance. If the reported results are valid, the paper would show that a YOLOv4-based architecture can be improved by about 6.2 AP with only a small speed penalty, which is practically meaningful for real-time detection. The paper does not appear to fit constants to a target result or rely on self-citations, and the reported ablation tables are internally monotonic. Its main weaknesses are that the scale-awareness mechanism is described in a way that is vacuous as written, the Recover layer is not defined, and all empirical claims rest on single runs without released code, weights, or error bars. The central architectural claim therefore needs substantial clarification and re-validation before the experimental attributions can be accepted.

major comments (3)
  1. [Section III-A, Eq. (2)] The scale-aware attention is vacuous as written. The text states that in the Concat layer the authors 'stack F1 with itself' because the neck has already performed feature interactions, yielding F ∈ R^{SL×H×W×C}. Since the slices along SL are identical copies of the same neck feature, Eq. (2) averages over S and C and produces attention weights that can only rescale identical content by per-slice scalars. Such weights cannot select among feature maps that actually differ in scale, so the 'scale-awareness' claim is not implementable in the sense described. This directly undermines the attributions in Table IV (+1.8 AP for the TDA-YOLO Module) and Table V (+1.0 AP50 for Scale Att.), which credit the gain to scale selection. If the authors intended to stack the genuinely different pyramid levels P3/P4/P5, the manuscript omits the necessary spatial resizing to a common H×W and the definition of how the slices are aligned; either way, Eq. (2) cannot implement real scale-aware attention as stated.
  2. [Section III-A, 'Recover' layer] The output of the adapted Dynamic Blocks is F ∈ R^{SL×S×C}, and the paper says the Recover layer reshapes it to F ∈ R^{C×H×W}, but no definition of the Recover operation is given. It is not specified whether this is a mean over SL, a concatenation followed by convolution, a learned projection, or something else. Since the number of elements in SL×S×C equals C×H×W only when SL=1, a mere reshape is not generally possible, so the module is not fully defined. This is a load-bearing omission because the TDA-YOLO Module's input and output dimensions must be specified for the architecture to be reproducible and for Eq. (1)-(4) to be implementable.
  3. [Section IV-A and Tables IV-VI] All experimental results are reported as single runs without error bars, standard deviations, or multiple seeds, and no code or trained weights are released. The headline ablation increments are small: +1.8 AP for the complete TDA-YOLO Module, +1.0 AP50 for scale attention, and +0.5 to +1.5 AP for several other components. In real-time detector training, such differences can be comparable to run-to-run variation under a fixed schedule, especially with heavy augmentation (Mosaic, Mixup, label smoothing). The claim that the TDA-YOLO Module contributes 'the largest gain' and the component-wise attributions in Tables V and VI therefore need either multiple seeds with variance estimates or public code/weights so that the results can be independently reproduced. Without this, the central empirical attribution is not yet established.
minor comments (6)
  1. [Throughout] There are numerous typographical errors, including 'samll-object' in the introduction, 'representations' in the abstract, 'primarly', 'paractices', 'schdule', 'representive', 'percise', and 'extraly'. These should be corrected in a thorough proofreading pass.
  2. [Section III-A] The dimension notation is inconsistent: the input is written F1 ∈ R^{SL×H×W×C}, but Eq. (2) writes F with dimensions SL×S×C without clearly defining the flattening/permutation order. Also, the hard sigmoid σ(x) = max(0, min(1, (x+1)/2)) is applied to a tensor in Eq. (2) but the elementwise nature is not stated explicitly.
  3. [Fig. 3] Figure 3 contains placeholder text '??, ??, ??, ??', which appears to be unfinished. The figure should be completed and its labels aligned with the equations.
  4. [Section IV-C, Table IV] The table reports parameter counts for each ablation, but the text in Section III-C claims a '0.4M reduction in parameters' from neck improvements while Table IV shows the neck step changing parameters from 65.5M to 65.1M. This is consistent, but the paper does not explicitly connect the claim to the table; adding a sentence would help.
  5. [Section I and Fig. 1] The speed/accuracy figure plots latency (ms) on the x-axis while the text and tables report FPS. The relationship is straightforward, but the figure should note the measurement condition (batch size 1 on V100) directly in the caption, as the reader would otherwise have to cross-reference the experimental setup.
  6. [Conclusions] The conclusion's admission that integration on more advanced baselines such as YOLOv7 is unexplored is appropriate and should remain, but the paper should also explicitly state the absence of code and weights, and ideally commit to releasing them if the results are to be independently verified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline AP gains are external benchmark measurements, the cited attention modules are prior external works, and no fitted parameter is renamed as a prediction.

full rationale

The paper's central claims are empirical: 3A-YOLO improves YOLOv4 by 6.2% AP on COCO test-dev (Table I), and the TDA-YOLO Module adds 1.8% AP in the Table IV ablation. These are measurements on held-out benchmarks, not quantities derived from the paper's own assumptions or fitted to the target result. The attention equations (Eqs. 1-4) are explicitly adapted from Dynamic Head [16], and the Coordinate Attention component is from [24]; neither is authored by the current team, so there is no load-bearing self-citation chain. The 'triple discriminative awareness' terminology is Dynamic Head's existing scale/spatial/task-aware decomposition, reused with attribution rather than renamed as an independent derivation. One internal consistency issue is worth noting but is not circularity: Section III-A says the Concat layer stacks 'F1 with itself,' which would make the SL axis of Eq. (2) contain identical copies and render genuine scale selection vacuous, and the Recover layer is left unspecified. If literal, this undermines the scale-awareness interpretation of the ablation gains, but it does not make the benchmark numbers an output of the paper's construction; the AP values remain independent observations. Because no reduction of a prediction to its input by equation, fitted parameter, or self-citation is exhibited, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its load-bearing elements are borrowed components and tuned design choices: the attention equations come from Dynamic Head and Dynamic ReLU, the positional module is Coordinate Attention, and the training recipe is standard. Free parameters are architectural and training hyperparameters, none of which are derived from first principles. The main unverified assumption is that single-run benchmark gains reflect robust causal contributions.

free parameters (4)
  • CA channel reduction ratio r = 16
    Equation (9) sets r=16. Table VI compares r=16 versus r=32 and reports 87.3 versus 86.3 AP50, so the choice was tuned on the VOC benchmark and directly affects the claimed CA contribution.
  • Number of adapted Dynamic Blocks = 2 for full models, 1 for Tiny
    Section III-A and Section III-D set this architecture choice, but no ablation is provided to justify 2 blocks over other counts. The central head design depends on this choice.
  • Training schedule (epochs, learning rate, batch size) = 300 epochs, initial lr 0.01, batch 256
    Section IV-A specifies these values as inherited from YOLOv5 and YOLOv4 practices. Reported AP depends on this schedule, which is not independently derived or varied.
  • Focal loss alpha balancing coefficient = not stated
    Section III-C says an alpha-balanced variant of Focal Loss is applied, but the alpha value is not specified. It is taken from RetinaNet and is not calibrated on the datasets used here.
assumptions (5)
  • standard math Hard sigmoid sigma(x)=max(0,min(1,(x+1)/2)) is an adequate gating approximation for attention weights.
    Used in Equation (2) and in the task-aware attention component, copied from Dynamic Head without new justification.
  • domain assumption Deformable convolution offsets and importance scalars are learnable and improve spatial attention on COCO and VOC.
    Section III-A, Equation (3), assumes gradient-based training learns useful offsets and masks even without the original Dynamic Head training setup.
  • domain assumption COCO and VOC ground-truth labels and AP metrics are reliable measures of detector quality.
    Section IV-A establishes evaluation on these datasets; all improvement claims rest entirely on these labels and metrics.
  • ad hoc to paper Stacking F1 with itself after the neck is a valid way to provide multi-scale input to the detection head.
    Section III-A first paragraph says the neck has done enough feature interaction and then stacks F1 with itself, but no evidence is given for why self-stacking is necessary or sufficient.
  • ad hoc to paper Placing Coordinate Attention on C3, C4, and C5 improves small-object capture and long-range semantics as stated.
    Section III-B and Figure 2 assert per-scale benefits, but the paper provides only a global ablation, not per-scale analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3A-YOLO: New Real-Time Object Detectors with Triple Discriminative Awareness and Coordinated Representations." pith.science (2026). https://pith.science/paper/K4FU5RS6

@misc{pith2026241207168,
  author       = {Pith},
  title        = {Pith review of: 3A-YOLO: New Real-Time Object Detectors with Triple Discriminative Awareness and Coordinated Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K4FU5RS6}},
  note         = {Machine review of arXiv:2412.07168}
}
read the original abstract

Recent research on real-time object detectors (e.g., YOLO series) has demonstrated the effectiveness of attention mechanisms for elevating model performance. Nevertheless, existing methods neglect to unifiedly deploy hierarchical attention mechanisms to construct a more discriminative YOLO head which is enriched with more useful intermediate features. To tackle this gap, this work aims to leverage multiple attention mechanisms to hierarchically enhance the triple discriminative awareness of the YOLO detection head and complementarily learn the coordinated intermediate representations, resulting in a new series detectors denoted 3A-YOLO. Specifically, we first propose a new head denoted TDA-YOLO Module, which unifiedly enhance the representations learning of scale-awareness, spatial-awareness, and task-awareness. Secondly, we steer the intermediate features to coordinately learn the inter-channel relationships and precise positional information. Finally, we perform neck network improvements followed by introducing various tricks to boost the adaptability of 3A-YOLO. Extensive experiments across COCO and VOC benchmarks indicate the effectiveness of our detectors.

Figures

Figures reproduced from arXiv: 2412.07168 by the authors.

Figure 1
Figure 1. Speed and accuracy trade-off of our 3A-YOLO series detectors and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The illustration of our 3A-YOLO. SPP is the Spatial Pyramid Pooling Layer [ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The illustration of our proposed TDA-YOLO Module. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The overall structure of our 3A-YOLO-Tiny. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Feature map visualizations coupled with the predictions of 3A-YOLO. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 34 canonical work pages

  1. [1]

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

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” IEEE T-PAMI, vol. 39, no. 6, pp. 1137–1149, 2016

    work page 2016

  2. [2]

    Jocher, “yolov5,” https://github.com/ultralytics/yolov5

    work page

  3. [3]

    Ssd: Single shot multibox detector,

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg, “Ssd: Single shot multibox detector,” in ECCV. Springer, 2016, pp. 21–37

    work page 2016

  4. [4]

    Yolov3: An incremen- tal improvement,

    Ali Farhadi and Joseph Redmon, “Yolov3: An incremen- tal improvement,” in CVPR, 2018, vol. 1804, pp. 1–6

    work page 2018

  5. [5]

    Ef- ficientdet: Scalable and efficient object detection,

    Mingxing Tan, Ruoming Pang, and Quoc V Le, “Ef- ficientdet: Scalable and efficient object detection,” in CVPR, 2020, pp. 10781–10790

    work page 2020

  6. [6]

    Yolov4: Optimal speed and accuracy of object detection,

    Alexey Bochkovskiy, Chien-Yao Wang, and Hong- Yuan Mark Liao, “Yolov4: Optimal speed and accuracy of object detection,” arXiv preprint arXiv:2004.10934 , 2020

    arXiv 2004

  7. [7]

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

    Chien-Yao Wang, Alexey Bochkovskiy, and Hong- Yuan Mark Liao, “Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” in CVPR, 2023, pp. 7464–7475

    work page 2023

  8. [8]

    Squeeze-and-excitation networks,

    Jie Hu, Li Shen, and Gang Sun, “Squeeze-and-excitation networks,” in CVPR, 2018, pp. 7132–7141

    work page 2018

Show all 40 references
  1. [9]

    Eca-net: Efficient channel attention for deep convolutional neural net- works,

    Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu, “Eca-net: Efficient channel attention for deep convolutional neural net- works,” in CVPR, 2020, pp. 11534–11542

    2020

  2. [10]

    Ebse-yolo: high precision recognition algorithm for small target foreign object detection,

    Shuohe Wang, Yongda Wang, Yujian Chang, et al., “Ebse-yolo: high precision recognition algorithm for small target foreign object detection,” IEEE Access, vol. 11, pp. 57951–57964, 2023

    2023

  3. [11]

    Yolo-se: Improved yolov8 for remote sensing object detection and recog- nition,

    Tianyong Wu and Youkou Dong, “Yolo-se: Improved yolov8 for remote sensing object detection and recog- nition,” Applied Sciences , vol. 13, no. 24, pp. 12977, 2023

    2023

  4. [12]

    Sb- myv3: improved mobyolov3 a bam attention-based ap- proach for obscene image and video detection,

    Sonali Samal, Yu-Dong Zhang, Thippa Reddy Gadekallu, Rajashree Nayak, and Bunil Kumar Balabantaray, “Sb- myv3: improved mobyolov3 a bam attention-based ap- proach for obscene image and video detection,” Expert Systems, vol. 40, no. 6, pp. e13230, 2023

    2023

  5. [13]

    A simple and light-weight attention module for convolutional neural networks,

    Jongchan Park, Sanghyun Woo, Joon-Young Lee, and In So Kweon, “A simple and light-weight attention module for convolutional neural networks,” IJCV, vol. 128, no. 4, pp. 783–798, 2020

    2020

  6. [14]

    Cbam: Convo- lutional block attention module,

    Sanghyun Woo, Jongchan Park, et al., “Cbam: Convo- lutional block attention module,” in ECCV, 2018, pp. 3–19

    2018

  7. [15]

    Yolo- sld: An attention mechanism-improved yolo for license plate detection,

    Ming-An Chung, Yu-Jou Lin, and Chia-Wei Lin, “Yolo- sld: An attention mechanism-improved yolo for license plate detection,” IEEE Access , 2024

    2024

  8. [16]

    Dynamic head: Unifying object detection heads with attentions,

    Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang, “Dynamic head: Unifying object detection heads with attentions,” in CVPR, 2021, pp. 7373–7382

    2021

  9. [17]

    Microsoft coco: Common objects in context,

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, et al., “Microsoft coco: Common objects in context,” in ECCV. Springer, 2014, pp. 740–755

    2014

  10. [18]

    The pascal visual object classes (voc) challenge,

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman, “The pascal visual object classes (voc) challenge,” IJCV, vol. 88, pp. 303–338, 2010

    2010

  11. [19]

    Focal loss for dense object detection,

    T-YLPG Ross and GKHP Doll ´ar, “Focal loss for dense object detection,” in CVPR, 2017, pp. 2980–2988

    2017

  12. [20]

    Deep residual learning for image recognition,

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778

    2016

  13. [21]

    Csp- net: A new backbone that can enhance learning capability of cnn,

    Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh, “Csp- net: A new backbone that can enhance learning capability of cnn,” in CVPR workshops , 2020, pp. 390–391

    2020

  14. [22]

    Parallel feature pyramid network for object detection,

    Seung-Wook Kim, Hyong-Keun Kook, Jee-Young Sun, Mun-Cheon Kang, and Sung-Jea Ko, “Parallel feature pyramid network for object detection,” in ECCV, 2018, pp. 234–250

    2018

  15. [23]

    Path aggregation network for instance segmentation,

    Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia, “Path aggregation network for instance segmentation,” in CVPR, 2018, pp. 8759–8768

    2018

  16. [24]

    Coordinate attention for efficient mobile network design,

    Qibin Hou, Daquan Zhou, and Jiashi Feng, “Coordinate attention for efficient mobile network design,” in CVPR, 2021, pp. 13713–13722

    2021

  17. [25]

    Learning spatial attention for face super-resolution,

    Chaofeng Chen, Dihong Gong, Hao Wang, Zhifeng Li, and Kwan-Yee K Wong, “Learning spatial attention for face super-resolution,” IEEE TIP , vol. 30, pp. 1219– 1231, 2020

    2020

  18. [26]

    Gather-excite: Exploiting feature context in convolutional neural networks,

    Jie Hu, Li Shen, Samuel Albanie, Gang Sun, and Andrea Vedaldi, “Gather-excite: Exploiting feature context in convolutional neural networks,” NeurIPS, vol. 31, 2018

    2018

  19. [27]

    Learning what and where to attend,

    Drew Linsley, Dan Shiebler, Sven Eberhardt, and Thomas Serre, “Learning what and where to attend,” arXiv preprint arXiv:1805.08819, 2018

    2018 arXiv

  20. [28]

    Spatial pyramid pooling in deep convolutional networks for visual recognition,

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Spatial pyramid pooling in deep convolutional networks for visual recognition,” IEEE T-PAMI, vol. 37, no. 9, pp. 1904–1916, 2015

    1904

  21. [29]

    Scaled-yolov4: Scaling cross stage partial network,

    Chien-Yao Wang, Alexey Bochkovskiy, and Hong- Yuan Mark Liao, “Scaled-yolov4: Scaling cross stage partial network,” in CVPR, 2021, pp. 13029–13038

    2021

  22. [30]

    Dynamic relu,

    Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, et al., “Dynamic relu,” in ECCV. Springer, 2020, pp. 351–367

    2020

  23. [31]

    Distance-iou loss: Faster and better learning for bounding box regression,

    Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rong- guang Ye, and Dongwei Ren, “Distance-iou loss: Faster and better learning for bounding box regression,” in AAAI, 2020, vol. 34, pp. 12993–13000

    2020

  24. [32]

    You only look one- level feature,

    Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun, “You only look one- level feature,” in CVPR, 2021, pp. 13039–13048

    2021

  25. [33]

    Pp-yolo: An effective and effi- cient implementation of object detector,

    Xiang Long et al., “Pp-yolo: An effective and effi- cient implementation of object detector,” arXiv preprint arXiv:2007.12099, 2020

    2007 arXiv

  26. [34]

    Learning spatial fusion for single-shot object detection,

    Songtao Liu, Di Huang, and Yunhong Wang, “Learning spatial fusion for single-shot object detection,” arXiv preprint arXiv:1911.09516, 2019

    1911 arXiv

  27. [35]

    Pp-yolov2: A practical object detector,

    Xin Huang, Xinxin Wang, Wenyu Lv, Xiaying Bai, Xiang Long, Kaipeng Deng, Qingqing Dang, Shumin Han, Qiwen Liu, Xiaoguang Hu, et al., “Pp-yolov2: A practical object detector,” arXiv preprint arXiv:2104.10419 , 2021

    2021 arXiv

  28. [36]

    Hierarchical shot detector,

    Jiale Cao, Yanwei Pang, Jungong Han, and Xuelong Li, “Hierarchical shot detector,” in ICCV, 2019, pp. 9705– 9714

    2019

  29. [37]

    Blitznet: A real-time deep network for scene understanding,

    Nikita Dvornik, Konstantin Shmelkov, Julien Mairal, and Cordelia Schmid, “Blitznet: A real-time deep network for scene understanding,” in ICCV, 2017, pp. 4154–4162

    2017

  30. [38]

    Self-knowledge distillation with progressive refinement of targets,

    Kyungyul Kim, ByeongMoon Ji, Doyoung Yoon, and Sangheum Hwang, “Self-knowledge distillation with progressive refinement of targets,” in ICCV, 2021, pp. 6567–6576

    2021

  31. [39]

    Couplenet: Coupling global structure with local parts for object detection,

    Yousong Zhu, Chaoyang Zhao, Jinqiao Wang, Xu Zhao, Yi Wu, and Hanqing Lu, “Couplenet: Coupling global structure with local parts for object detection,” in ICCV, 2017, pp. 4126–4134

    2017

  32. [40]

    Nanodet-plus: Super fast and high ac- curacy lightweight anchor-free object detection model,

    Rangi Lyu, “Nanodet-plus: Super fast and high ac- curacy lightweight anchor-free object detection model,” https://github. com/RangiLyu/nanodet, 2021

    2021

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.