Back

YOLOX: Exceeding YOLO Series in 2021

Zheng GeSongtao LiuFeng WangZeming LiJian Sun

article2021arXiv6,198 citations

Introduces an anchor-free redesign of the YOLO object detection architecture featuring a decoupled head and SimOTA dynamic label assignment, surpassing existing benchmarks in accuracy and inference speed across lightweight and large-scale models.

Listen

Real-time computer vision systems require a balance between detection accuracy and execution speed. Although recent academic advances introduced anchor-free architectures and advanced label assignment strategies to enhance performance, mainstream industrial detectors such as the YOLO family remained reliant on anchor-based pipelines with hand-crafted training rules. The article set out to modernize these widely deployed detectors by incorporating anchor-free mechanisms, decoupled detection heads, and dynamic label assignment into a new high-performance system named YOLOX.

The developers evaluated this framework across multiple model sizes using standard benchmark datasets. They modernized the base architecture by separating the classification and localization branches, removing predefined anchor boxes in favor of direct coordinate prediction, and implementing a simplified dynamic label assignment method called SimOTA. The system was validated against baseline architectures across various hardware platforms and edge-device constraints, maintaining standardized training schedules and evaluation protocols.

The findings show substantial improvements in both accuracy and efficiency across all model scales. Upgrading the baseline architecture boosted standard benchmark accuracy from 38.5% to 47.3%, outperforming previous best practices by 3.0 percentage points. For high-capacity models, the design achieved 50.0% accuracy at 68.9 frames per second, exceeding comparable state-of-the-art baselines by 1.8 percentage points. In lightweight and mobile configurations, the ultra-compact version achieved 25.3% accuracy with only 0.91 million parameters, surpassing alternative lightweight detectors while reducing computational requirements.

These results demonstrate that anchor-free designs and advanced assignment strategies can improve accuracy without incurring meaningful latency penalties. Removing anchor mechanisms significantly reduces heuristic tuning and cross-device transmission bottlenecks, making the models cheaper to deploy and easier to maintain across production environments ranging from edge devices to enterprise servers.

Organizations deploying computer vision should consider adopting this modernized architecture, selecting the model scale that fits their specific hardware constraints. Data augmentation should be calibrated carefully, using strong image mixing for large models and reduced distortion for compact networks. Future work should evaluate these techniques on larger transformer-based architectures and emerging multi-scale frameworks.

Megvii Technologyautonomous drivingcommunicating sequential processesdetection algorithmdense object detectionanchor-free detectorsYOLO series
arXiv: 2107.08430Megvii-BaseDetection/YOLOX
Cover for YOLOX: Exceeding YOLO Series in 2021

Abstract

In this report, we present some experienced improvements to YOLO series, forming a new high-performance detector -- YOLOX. We switch the YOLO detector to an anchor-free manner and conduct other advanced detection techniques, i.e., a decoupled head and the leading label assignment strategy SimOTA to achieve state-of-the-art results across a large scale range of models: For YOLO-Nano with only 0.91M parameters and 1.08G FLOPs, we get 25.3% AP on COCO, surpassing NanoDet by 1.8% AP; for YOLOv3, one of the most widely used detectors in industry, we boost it to 47.3% AP on COCO, outperforming the current best practice by 3.0% AP; for YOLOX-L with roughly the same amount of parameters as YOLOv4-CSP, YOLOv5-L, we achieve 50.0% AP on COCO at a speed of 68.9 FPS on Tesla V100, exceeding YOLOv5-L by 1.8% AP. Further, we won the 1st Place on Streaming Perception Challenge (Workshop on Autonomous Driving at CVPR 2021) using a single YOLOX-L model. We hope this report can provide useful experience for developers and researchers in practical scenes, and we also provide deploy versions with ONNX, TensorRT, NCNN, and Openvino supported. Source code is at this https URL.

Topics

autonomous drivingcommunicating sequential processesdetection algorithmsource codesaccess points

Keywords

dense object detectionanchor-free detectorsYOLO seriesstrong data augmentationinference speedinstance mask annotationsstreaming perception

Table of Contents

  • 1 Introduction
  • 2 YOLOX
  • 2.1 YOLOX-DarkNet53
  • 2.2 Other Backbones
  • 3 Comparison with the SOTA
  • 4 1st Place on Streaming Perception Challenge (WAD at CVPR 2021)
  • 5 Conclusion
  • References

Knowls

  1. Knowl 1 — Lite Decoupled Head Architecture for YOLO Detectors

    model/method

    Traditional YOLO detectors (e.g., YOLOv3 through YOLOv5) utilize a coupled detection head where classification categories, bounding box regression offsets, and objectness confidence scores are predicted jointly from shared feature channels at each feature pyramid level (P3,P4,P5P_3, P_4, P_5).

    YOLOX decouples classification and localization into separate branches to mitigate the task conflict between classification and bounding box regression:

    1. For each feature pyramid level with input channels CinC_{in}, a 1×11 \times 1 convolution reduces the feature channel dimension to 256.
    2. The 256-channel feature map is fed into two parallel branches, each consisting of two consecutive 3×33 \times 3 convolutional layers.
    3. The classification branch outputs class logits of shape H×WimesCclsH \times W imes C_{cls}, where HH and WW denote spatial dimensions and CclsC_{cls} is the number of object classes.
    4. The regression branch forks into two sub-outputs: a bounding box regression map of shape H×W×4H \times W \times 4 and an IoU/objectness prediction map of shape H×W×1H \times W \times 1.

    On a DarkNet53 baseline at 640×640640 \times 640 resolution, replacing the coupled head with the lite decoupled head increases COCO AP from 38.5% to 39.6%, accelerates training convergence, and adds 1.1 ms inference latency (11.6 ms vs. 10.5 ms) on a Tesla V100 with batch size 1 and FP16 precision.

  2. Knowl 2 — SimOTA Dynamic Label Assignment Algorithm

    algorithm

    SimOTA (Simplified Optimal Transport Assignment) is an anchor/grid assignment strategy that dynamically assigns positive predictions to ground-truth objects based on prediction-to-ground-truth matching costs without requiring iterative Sinkhorn-Knopp solvers.

    Input: Ground-truth bounding boxes {gi}i=1M\{g_i\}_{i=1}^M, predicted bounding boxes and class logits {pj}j=1N\{p_j\}_{j=1}^N, loss balancing parameter λ\lambda, center region radius rr.
    Output: Label assignment mapping positive predictions to corresponding ground-truth objects.
    for each ground-truth gig_i from 11 to MM do
        Identify candidate predictions Pi⊂{pj}j=1NP_i \subset \{p_j\}_{j=1}^N located within a fixed spatial center region of gig_i
        for each candidate prediction pj∈Pip_j \in P_i do
            Compute classification loss LijclsL_{ij}^{\text{cls}} (e.g., Binary Cross Entropy)
            Compute bounding box regression loss LijregL_{ij}^{\text{reg}} (e.g., IoU loss)
            Compute matching cost cij=Lijcls+λLijregc_{ij} = L_{ij}^{\text{cls}} + \lambda L_{ij}^{\text{reg}}
        end for
        Estimate dynamic count kik_i by summing the IoU values of the top candidate predictions for gig_i (rounded to an integer)
        Select the kik_i predictions from PiP_i with the smallest cost values cijc_{ij}
        Assign these kik_i predictions as positive samples for gig_i
    end for
    Assign all unassigned prediction grids as negative (background) samples
    return assignment mapping

    SimOTA avoids additional solver hyperparameters and excessive compute overhead from optimal transport solvers while improving YOLOX-DarkNet53 performance on COCO from 45.0% AP to 47.3% AP (+2.3% AP).

  3. Knowl 3 — Anchor-Free Representation and Multi-Positive Center Sampling

    model/method

    YOLOX converts the anchor-based detection mechanism of traditional YOLO models into an anchor-free design:

    1. Anchor-Free Prediction: Rather than predicting 3 anchor boxes per grid location, the detector outputs exactly 1 prediction per spatial location (x,y)(x, y) on each feature pyramid level. The network directly predicts 4 coordinate values: the 2 offsets relative to the top-left corner of the grid cell, and the height and width of the predicted bounding box.
    2. Feature Level Assignment: A pre-defined object scale range designates which feature pyramid level is assigned to each ground truth.
    3. Multi-Positive Center Sampling: Single-center assignment designates only the single grid cell containing the ground-truth center as positive, discarding nearby high-quality predictions and exacerbating positive/negative sample imbalance. YOLOX defines the entire 3×33 \times 3 grid neighborhood surrounding the ground-truth center as positive candidates.

    On the YOLOX-DarkNet53 baseline at 640×640640 \times 640 resolution on COCO, switching from anchor-based to single-center anchor-free prediction improves AP from 42.0% to 42.9% while reducing model parameters from 63.86M to 63.72M and GFLOPs from 186.0 to 185.3. Adding multi-positive center sampling further increases AP from 42.9% to 45.0%.

  4. Knowl 4 — Model-Scale-Aware Data Augmentation Policy

    model/method

    YOLOX applies a strong data augmentation policy comprising Mosaic and MixUp with scale jittering, adjusted according to model capacity, and disabled during late-stage training:

    1. MixUp with Scale Jittering: Before blending two images via MixUp, both images are randomly scaled by a jitter factor. This acts as an effective substitute for CopyPaste augmentation without requiring instance mask annotations (achieving 49.5% AP on YOLOX-L compared to 49.4% AP with CopyPaste).
    2. Scale-Dependent Augmentation Tuning: Strong augmentations benefit large models but impair small/compact models. For large models (e.g., YOLOX-L), standard Mosaic scale jittering in the range [0.1,2.0][0.1, 2.0] and MixUp boost AP by +0.9% (48.6% to 49.5%). For compact models (e.g., YOLOX-Nano, YOLOX-Tiny, YOLOX-S), MixUp is removed and Mosaic scale jittering is narrowed to [0.5,1.5][0.5, 1.5], which improves YOLOX-Nano AP from 24.0% to 25.3%.
    3. Augmentation Shutdown: Mosaic and MixUp augmentations are disabled for the final 15 epochs of the 300-epoch training schedule, allowing the network to settle and optimize bounding box predictions on uncorrupted, natural image distributions.

  5. Knowl 5 — Roadmap of Cumulative Improvements for YOLOX-DarkNet53

    data/table

    The ablation experiments illustrate the cumulative impact of each architectural and training enhancement applied to the YOLOv3-SPP baseline on the COCO validation set. All models are evaluated at 640×640640 \times 640 resolution with FP16 precision and batch size 1 on a Tesla V100 (latency and FPS measured without post-processing).

    Methods AP (%) Parameters GFLOPs Latency FPS
    YOLOv3-ultralytics 44.3 63.00 M 157.3 10.5 ms 95.2
    YOLOv3 baseline 38.5 63.00 M 157.3 10.5 ms 95.2
    + decoupled head 39.6 (+1.1) 63.86 M 186.0 11.6 ms 86.2
    + strong augmentation 42.0 (+2.4) 63.86 M 186.0 11.6 ms 86.2
    + anchor-free 42.9 (+0.9) 63.72 M 185.3 11.1 ms 90.1
    + multi positives 45.0 (+2.1) 63.72 M 185.3 11.1 ms 90.1
    + SimOTA 47.3 (+2.3) 63.72 M 185.3 11.1 ms 90.1
    + NMS free (optional) 46.5 (-0.8) 67.27 M 205.1 13.5 ms 74.1

    The combined additions (decoupled head, strong augmentation, anchor-free design, multi-positive assignment, and SimOTA) advance the baseline detector from 38.5% AP to 47.3% AP, exceeding YOLOv3-ultralytics by 3.0% AP.

  6. Knowl 6 — Comparative Performance of YOLOX vs. YOLOv5 across Standard Model Scales

    data/table

    Adopting the YOLOv5 modified CSPNet backbone, SiLU activations, and PAN neck under identical scaling rules produces four standard models: YOLOX-S, YOLOX-M, YOLOX-L, and YOLOX-X. Performance is evaluated on COCO val at 640×640640 \times 640 resolution, FP16 precision, and batch size 1 on a Tesla V100.

    Models AP (%) Parameters GFLOPs Latency
    YOLOv5-S 36.7 7.3 M 17.1 8.7 ms
    YOLOX-S 39.6 (+2.9) 9.0 M 26.8 9.8 ms
    YOLOv5-M 44.5 21.4 M 51.4 11.1 ms
    YOLOX-M 46.4 (+1.9) 25.3 M 73.8 12.3 ms
    YOLOv5-L 48.2 47.1 M 115.6 13.7 ms
    YOLOX-L 50.0 (+1.8) 54.2 M 155.6 14.5 ms
    YOLOv5-X 50.4 87.8 M 219.0 16.0 ms
    YOLOX-X 51.2 (+0.8) 99.1 M 281.9 17.3 ms

    Across all model sizes from S to X, YOLOX consistently achieves between +0.8% and +2.9% higher AP than the corresponding YOLOv5 models with marginal latency increments from the decoupled head.

  7. Knowl 7 — Benchmark of Lightweight Detectors: YOLOX-Tiny and YOLOX-Nano

    data/table

    For lightweight and mobile deployment, YOLOX is scaled down to YOLOX-Tiny and YOLOX-Nano (which incorporates depthwise separable convolutions). Evaluated at 416×416416 \times 416 resolution on COCO val:

    Models AP (%) Parameters GFLOPs
    YOLOv4-Tiny 21.7 6.06 M 6.96
    PPYOLO-Tiny 22.7 4.20 M –
    YOLOX-Tiny 32.8 (+10.1) 5.06 M 6.45
    NanoDet 23.5 0.95 M 1.20
    YOLOX-Nano 25.3 (+1.8) 0.91 M 1.08

    YOLOX-Tiny exceeds YOLOv4-Tiny by 10.1% AP with fewer parameters and FLOPs. YOLOX-Nano achieves 25.3% AP with 0.91M parameters and 1.08 GFLOPs, surpassing NanoDet by 1.8% AP using a smaller parameter footprint.

  8. Knowl 8 — COCO 2017 Test-Dev Benchmark Comparison

    data/table

    Comparison of YOLOX models against state-of-the-art detectors on COCO 2017 test-dev. All models were trained for 300 epochs and tested on a Tesla V100 GPU.

    Method Backbone Size FPS AP (%) AP50\text{AP}_{50} AP75\text{AP}_{75} APS\text{AP}_S APM\text{AP}_M APL\text{AP}_L
    YOLOv3 + ASFF* Darknet-53 608 45.5 42.4 63.0 47.4 25.5 45.7 52.3
    YOLOv3 + ASFF* Darknet-53 800 29.4 43.9 64.1 49.2 27.0 46.6 53.4
    EfficientDet-D0 Efficient-B0 512 98.0 33.8 52.2 35.8 12.0 38.3 51.2
    EfficientDet-D1 Efficient-B1 640 74.1 39.6 58.6 42.3 17.9 44.3 56.0
    EfficientDet-D2 Efficient-B2 768 56.5 43.0 62.3 46.2 22.5 47.0 58.4
    EfficientDet-D3 Efficient-B3 896 34.5 45.8 65.0 49.3 26.6 49.4 59.8
    PP-YOLOv2 ResNet50-vd-dcn 640 68.9 49.5 68.2 54.4 30.7 52.9 61.2
    PP-YOLOv2 ResNet101-vd-dcn 640 50.3 50.3 69.0 55.3 31.6 53.9 62.4
    YOLOv4 CSPDarknet-53 608 62.0 43.5 65.7 47.3 26.7 46.7 53.3
    YOLOv4-CSP Modified CSP 640 73.0 47.5 66.2 51.7 28.2 51.2 59.8
    YOLOv3-ultralytics Darknet-53 640 95.2 44.3 64.6 – – – –
    YOLOv5-M Modified CSP v5 640 90.1 44.5 63.1 – – – –
    YOLOv5-L Modified CSP v5 640 73.0 48.2 66.9 – – – –
    YOLOv5-X Modified CSP v5 640 62.5 50.4 68.8 – – – –
    YOLOX-DarkNet53 Darknet-53 640 90.1 47.4 67.3 52.1 27.5 51.5 60.9
    YOLOX-M Modified CSP v5 640 81.3 46.4 65.4 50.6 26.3 51.0 59.9
    YOLOX-L Modified CSP v5 640 69.0 50.0 68.5 54.5 29.8 54.5 64.4
    YOLOX-X Modified CSP v5 640 57.8 51.2 69.6 55.7 31.2 56.1 66.1

    YOLOX-L achieves 50.0% AP at 69.0 FPS on a single Tesla V100, outperforming YOLOv4-CSP (47.5% AP at 73.0 FPS) and YOLOv5-L (48.2% AP at 73.0 FPS). YOLOX-X achieves 51.2% AP at 57.8 FPS.

  9. Knowl 9 — Optional End-to-End NMS-Free Detection Module

    model/method

    YOLOX supports an optional end-to-end (NMS-free) configuration following Zhou et al. by introducing two additional convolutional layers, one-to-one label assignment during auxiliary matching, and a stop-gradient operation.

    While a coupled detection head experiences a substantial drop of 4.2% AP when switching from NMS post-processing to end-to-end training (38.5% down to 34.3% AP), the decoupled head limits this degradation to 0.8% AP (39.6% down to 38.8% AP on vanilla YOLO, and 47.3% down to 46.5% AP on the full YOLOX-DarkNet53).

    Because end-to-end mode incurs a modest accuracy penalty (-0.8% AP) and reduces inference speed from 90.1 FPS (11.1 ms) to 74.1 FPS (13.5 ms), it is retained as an optional feature rather than enabled in default YOLOX models.

  10. Knowl 10 — Streaming Perception Strategy Using YOLOX-L

    empirical result

    Streaming perception jointly evaluates detection accuracy and latency on continuous sensor streams (e.g., at 30 FPS) via streaming accuracy, forcing algorithms to account for environmental state changes that occur while computation is running.

    For a 30 FPS input stream, the optimal operating point corresponds to a maximum latency threshold of ≤33\le 33 ms per frame. A YOLOX-L model optimized with TensorRT was deployed within this ≤33\le 33 ms budget, maintaining real-time alignment with the 30 FPS stream and securing 1st place in the CVPR 2021 Workshop on Autonomous Driving (WAD) Streaming Perception Challenge.

Coverage note — None was omitted; all key technical innovations, algorithmic details, ablation studies, multi-scale model comparisons, lightweight detector architectures, and competitive benchmark results have been captured.

References

  1. 1.Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 1, 2, 3, 6
  2. 2.Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 4
  3. 3.Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In CVPR, 2021. 3
  4. 4.Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In CVPR, 2021. 1, 4
  5. 5.Zheng Ge, Jianfeng Wang, Xin Huang, Songtao Liu, and Osamu Yoshie. Lla: Loss-aware label assignment for dense pedestrian detection. arXiv preprint arXiv:2101.04307, 2021. 4
  6. 6.Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 5
  7. 7.glenn jocher et al. yolov5. https://github.com/ultralytics/yolov5, 2021. 1, 2, 3, 5, 6
  8. 8.Priya Goyal, Piotr Doll'ar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 2
  9. 9.Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1
  10. 10.Zhang Hongyi, Cisse Moustapha, N. Dauphin Yann, and David Lopez-Paz. mixup: Beyond empirical risk minimization. ICLR, 2018. 3
  11. 11.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. 3, 6
  12. 12.Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, 2020. 1, 4
  13. 13.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. 2
  14. 14.Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, 2018. 1, 3
  15. 15.Mengtian Li, Yuxiong Wang, and Deva Ramanan. Towards streaming perception. In ECCV, 2020. 5, 6
  16. 16.Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll'ar. Focal loss for dense object detection. In ICCV, 2017. 2
  17. 17.Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll'ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2
  18. 18.Songtao Liu, Di Huang, and Yunhong Wang. Learning spatial fusion for single-shot object detection. arXiv preprint arXiv:1911.09516, 2019. 6
  19. 19.Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, 2018. 2, 5
  20. 20.Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance segmentation. In CVPR, 2018. 2
  21. 21.Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 5
  22. 22.Yuchen Ma, Songtao Liu, Zeming Li, and Jian Sun. Iqdet: Instance-wise quality distribution sampling for object detection. In CVPR, 2021. 1, 4
  23. 23.Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 1
  24. 24.Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In CVPR, 2017. 1, 3
  25. 25.Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 1, 2, 3
  26. 26.Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1
  27. 27.Guanglu Song, Yu Liu, and Xiaogang Wang. Revisiting the sibling head in object detector. In CVPR, 2020. 2
  28. 28.Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In CVPR, 2020. 6
  29. 29.Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 1, 2, 3, 4
  30. 30.Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Scaled-yolov4: Scaling cross stage partial network. arXiv preprint arXiv:2011.08036, 2020. 1, 5, 6
  31. 31.Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh. Cspnet: A new backbone that can enhance learning capability of cnn. In CVPR workshops, 2020. 2, 5
  32. 32.Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, and Nanning Zheng. End-to-end object detection with fully convolutional network. In CVPR, 2020. 1
  33. 33.Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, and Nanning Zheng. End-to-end object detection with fully convolutional network. In CVPR, 2021. 4
  34. 34.Yue Wu, Yinpeng Chen, Lu Yuan, Zicheng Liu, Lijuan Wang, Hongzhi Li, and Yun Fu. Rethinking classification and localization for object detection. In CVPR, 2020. 2
  35. 35.Yue Wu, Yinpeng Chen, Lu Yuan, Zicheng Liu, Lijuan Wang, Hongzhi Li, and Yun Fu. Rethinking classification and localization for object detection. In CVPR, 2020. 2
  36. 36.Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, 2020. 1, 4
  37. 37.Xiaosong Zhang, Fang Wan, Chang Liu, Rongrong Ji, and Qixiang Ye. Freeanchor: Learning to match anchors for visual object detection. In NeurIPS, 2019. 1, 4
  38. 38.Zhi Zhang, Tong He, Hang Zhang, Zhongyuan Zhang, Junyuan Xie, and Mu Li. Bag of freebies for training object detection neural networks. arXiv preprint arXiv:1902.04103, 2019. 3, 5
  39. 39.Qiang Zhou, Chaohui Yu, Chunhua Shen, Zhibin Wang, and Hao Li. Object detection made simpler by eliminating heuristic nms. arXiv preprint arXiv:2101.11782, 2021. 1, 4
  40. 40.Xingyi Zhou, Dequan Wang, and Philipp Krähenbühl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 1, 3
  41. 41.Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020. 1, 4

Citation

MLA
Ge, Z., et al. “YOLOX: Exceeding YOLO Series in 2021”. arXiv, 2021, https://doi.org/10.48550/arxiv.2107.08430.
APA
Ge, Z., Liu, S., Wang, F., Li, Z., & Sun, J. (2021). YOLOX: Exceeding YOLO Series in 2021. arXiv. https://doi.org/10.48550/arxiv.2107.08430
Chicago
Ge, Z., S. Liu, F. Wang, Z. Li, and J. Sun. 2021. “YOLOX: Exceeding YOLO Series in 2021”. Preprint, ArXiv. https://doi.org/10.48550/arxiv.2107.08430.
Harvard
Ge, Z. et al. (2021) “YOLOX: Exceeding YOLO Series in 2021”. arXiv. Available at: https://doi.org/10.48550/arxiv.2107.08430.
Vancouver
1. Ge Z, Liu S, Wang F, Li Z, Sun J (2021) YOLOX: Exceeding YOLO Series in 2021. https://doi.org/10.48550/arxiv.2107.08430

BibTeX

@misc{https://doi.org/10.48550/arxiv.2107.08430,
  doi = {10.48550/ARXIV.2107.08430},
  url = {https://arxiv.org/abs/2107.08430},
  author = {Ge, Zheng and Liu, Songtao and Wang, Feng and Li, Zeming and Sun, Jian},
  keywords = {Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {YOLOX: Exceeding YOLO Series in 2021},
  publisher = {arXiv},
  year = {2021},
  copyright = {arXiv.org perpetual, non-exclusive license}
}
Metadata:DOI registry

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Authors