Back

Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive Training Sample Selection

Shifeng ZhangCheng ChiYongqiang YaoZhen LeiStan Z. Li

article2019CVPR2,210 citationsBest Paper Nomination

Reveals that positive and negative sample selection is the essential difference between anchor-based and anchor-free object detectors, introducing an Adaptive Training Sample Selection (ATSS) strategy that unifies both paradigms and boosts detection accuracy without added overhead.

Listen

Modern computer vision systems rely heavily on object detection models to identify and locate items in images for applications like visual recognition and surveillance. For years, the field was divided into anchor-based systems, which predict object boundaries starting from preset reference boxes, and anchor-free systems, which detect objects starting from key points. While anchor-free systems recently demonstrated superior performance, the technical reasons driving this performance gap remained misunderstood.

The article set out to determine the fundamental root cause of the performance disparity between anchor-based and anchor-free detectors and to develop an automated method to bridge that gap.

To conduct a fair evaluation, the authors aligned implementation details—such as normalization and loss functions—between representative anchor-based (RetinaNet) and anchor-free (FCOS) frameworks using the standard MS COCO dataset of 80 object classes. By isolating and testing the remaining differences, the researchers investigated whether the regression starting point (box versus point) or the sample selection strategy (how training data is classified as positive or negative) drove model accuracy.

The findings reveal that the starting status—whether predicting from an anchor point or an anchor box—has virtually no effect on final accuracy (yielding identical 37.8% Average Precision scores under matched conditions). Instead, the essential driver of performance is how models define positive (foreground) and negative (background) training samples. In response, the authors introduced Adaptive Training Sample Selection (ATSS), an automated mechanism that calculates dynamic thresholds based on the mean and standard deviation of candidate sample overlap. When applied to standard detectors, ATSS improved RetinaNet by 2.3% and FCOS by 1.4% without adding computational overhead. Furthermore, experiments demonstrated that placing multiple preset anchor boxes per location becomes entirely unnecessary once sample selection is optimized. Combined with advanced backbones and testing strategies, ATSS achieved a state-of-the-art 50.7% Average Precision.

These results show that system designers do not need to choose between anchor-based and anchor-free designs based on perceived accuracy differences, nor do they need to waste computational resources managing complex, dense anchor grids. By eliminating sensitive manual tuning parameters (such as rigid overlap thresholds), ATSS simplifies detector design, lowers engineering maintenance, and boosts overall model accuracy at zero extra inference cost.

Development teams building or deploying vision systems should adopt adaptive sample selection strategies and streamline architectures to use single presets per location rather than dense multi-anchor grids. Organizations should also consider incorporating ATSS into existing computer vision pipelines to capture immediate performance gains. Future work should further explore the underlying role and utility of multiple anchors under adaptive frameworks.

Confidence in these findings is high given the strict experimental controls and validated benchmarks across multiple model architectures on the MS COCO dataset. However, stakeholders should note that the evaluations primarily focus on standard one-stage and center-based detection frameworks, meaning custom or atypical architectures should be independently piloted before broad implementation.

Aerospace Information Research Institute, Chinese Academy of SciencesBeijing University of Posts and TelecommunicationsInstitute of Automation, Chinese Academy of SciencesUniversity of Chinese Academy of SciencesWestlake Universitydetection algorithmgenetic selectionobject detectionAdaptive Training Sample Selectionanchor-free detectionanchor-based detector
arXiv: 1912.02424sfzhang15/ATSS
  • Paper: Focal Loss for Dense Object Detection, Tsung-Yi Lin et al. (2017). Reading Focal Loss for Dense Object Detection first is essential because the source paper builds directly upon RetinaNet's backbone and focal loss formulation to investigate training sample selection.
  • Paper: FCOS: Fully Convolutional One-Stage Object Detection, Zhi Tian et al. (2019). FCOS provides the foundational anchor-free detection framework whose positive and negative sample definitions are directly analyzed and compared against anchor-based methods in the source paper.
Cover for Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive Training Sample Selection

Abstract

Object detection has been dominated by anchor-based detectors for several years. Recently, anchor-free detectors have become popular due to the proposal of FPN and Focal Loss. In this paper, we first point out that the essential difference between anchor-based and anchor-free detection is actually how to define positive and negative training samples, which leads to the performance gap between them. If they adopt the same definition of positive and negative samples during training, there is no obvious difference in the final performance, no matter regressing from a box or a point. This shows that how to select positive and negative training samples is important for current object detectors. Then, we propose an Adaptive Training Sample Selection (ATSS) to automatically select positive and negative samples according to statistical characteristics of object. It significantly improves the performance of anchor-based and anchor-free detectors and bridges the gap between them. Finally, we discuss the necessity of tiling multiple anchors per location on the image to detect objects. Extensive experiments conducted on MS COCO support our aforementioned analysis and conclusions. With the newly introduced ATSS, we improve state-of-the-art detectors by a large margin to 50.7%50.7\% AP without introducing any overhead. The code is available at this https URL

Topics

detection algorithmgenetic selectionobject detectiontraining samplesupport vectorcorrelation analysistraining programs

Keywords

Adaptive Training Sample Selectionanchor-free detectionanchor-based detectoranchor boxesobject detectionnegative training samplesfeature pyramidmultiscale training

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 2.1 Anchor-based Detector
  • 2.2 Anchor-free Detector
  • 3 Difference Analysis of Anchor-based and Anchor-free Detection
  • 3.1 Experiment Setting
  • 3.2 Inconsistency Removal
  • 3.3 Essential Difference
  • 4 Adaptive Training Sample Selection
  • 4.1 Description
  • 4.2 Verification
  • 4.3 Analysis
  • 4.4 Comparison
  • 4.5 Discussion
  • 5 Conclusion
  • References

Knowls

  1. Knowl 1 — Adaptive Training Sample Selection Algorithm

    algorithm

    Adaptive Training Sample Selection (ATSS) automatically partitions preset candidate anchor boxes into positive and negative training samples based on the statistical distribution of Intersection over Union (IoU) scores for each ground-truth object. For each ground truth gg, candidate anchors are first gathered across all LL feature pyramid levels by selecting the kk anchors on each level whose center points have the smallest Euclidean distance to the center of gg. The mean mgm_g and standard deviation vgv_g of the IoU scores between these k×Lk \times L candidate anchors and gg are computed to define a dynamic IoU threshold tg=mg+vgt_g = m_g + v_g. Candidate anchors with IoU≥tg\text{IoU} \ge t_g whose spatial centers fall strictly inside gg are assigned as positive samples for gg; anchors not selected for any ground-truth object are negative samples.

    Input: Ground-truth boxes GG, number of feature pyramid levels LL, anchor sets AiA_i for each level i∈[1,L]i \in [1, L], all anchors A=⋃i=1LAiA = \bigcup_{i=1}^L A_i, candidate count k=9k = 9
    Output: Positive sample set PP, negative sample set NN
    P←∅P \leftarrow \emptyset
    for each ground-truth box g∈Gg \in G:
        Cg←∅C_g \leftarrow \emptyset
        for each level i∈[1,L]i \in [1, L]:
            Si←S_i \leftarrow select kk anchors from AiA_i whose center coordinates have the smallest L2L_2 distance to the center of gg
            Cg←Cg∪SiC_g \leftarrow C_g \cup S_i
        Dg←{IoU(c,g)∣c∈Cg}D_g \leftarrow \{\text{IoU}(c, g) \mid c \in C_g\}
        mg←Mean(Dg)m_g \leftarrow \text{Mean}(D_g)
        vg←Std(Dg)v_g \leftarrow \text{Std}(D_g)
        tg←mg+vgt_g \leftarrow m_g + v_g
        for each candidate c∈Cgc \in C_g:
            if IoU(c,g)≥tg\text{IoU}(c, g) \ge t_g and center of cc lies inside gg:
                P←P∪{c}P \leftarrow P \cup \{c\}
    for each anchor c∈Pc \in P assigned to multiple ground-truth boxes in GG:
        assign cc exclusively to the ground-truth box gg with which it has the highest IoU(c,g)\text{IoU}(c, g)
    N←A∖PN \leftarrow A \setminus P
    return P,NP, N

  2. Knowl 2 — Essential Difference Between Anchor-Based and Center-Based Anchor-Free Detectors

    empirical result

    When architectural differences (Group Normalization, Generalized IoU loss, centerness branches, ground-truth box constraints, and per-level scalar multipliers) are controlled between one-stage anchor-based detectors (e.g., RetinaNet with 1 anchor per location) and center-based anchor-free detectors (e.g., FCOS), two design differences remain:

    1. Classification Sample Selection: Defining positives and negatives via preset Intersection over Union (IoU) thresholds versus spatial bounding box inclusion paired with level-specific scale range intervals.
    2. Regression Starting Representation: Regressing bounding box coordinates from a 4-parameter preset anchor box (Δx,Δy,Δw,Δh)(\Delta x, \Delta y, \Delta w, \Delta h) versus regressing 4 boundary distances (l,t,r,b)(l, t, r, b) from a 2D anchor point.

    Evaluating these two orthogonal factors on the MS COCO validation dataset demonstrates that the regression starting status produces negligible performance differences under identical positive/negative sample definitions, whereas sample selection defines detector performance:

    Sample Selection Strategy Regression from Box Regression from Point
    Intersection over Union (IoU) 37.0% AP 36.9% AP
    Spatial and Scale Constraints 37.8% AP 37.8% AP

    Thus, the essential difference between one-stage anchor-based and center-based anchor-free detectors is how positive and negative training samples are defined, not whether regression starts from a point or a bounding box.

  3. Knowl 3 — Statistical Properties and Invariance of ATSS Dynamic Thresholding

    model/method

    The dynamic threshold tg=mg+vgt_g = m_g + v_g in Adaptive Training Sample Selection (ATSS) leverages statistical properties of object candidates to achieve scale-invariant sample assignment:

    1. Candidate Suitability via Mean (mgm_g): The mean IoU mg=1∣Cg∣∑c∈CgIoU(c,g)m_g = \frac{1}{|C_g|} \sum_{c \in C_g} \text{IoU}(c, g) measures the quality of the closest preset anchors. A high mgm_g indicates well-aligned anchor candidates, raising the threshold to select high-quality positives; a low mgm_g lowers the threshold to retain viable positives for harder objects.
    2. Pyramid Level Selectivity via Standard Deviation (vgv_g): The standard deviation vg=1∣Cg∣∑c∈Cg(IoU(c,g)−mg)2v_g = \sqrt{\frac{1}{|C_g|} \sum_{c \in C_g} (\text{IoU}(c, g) - m_g)^2} measures variance across feature pyramid levels. A high vgv_g signifies that a single pyramid level provides superior candidate fit over others, elevating tgt_g to restrict positive sample assignment to that specific level. A low vgv_g indicates multiple pyramid levels are suitable, yielding a lower threshold that samples across several levels.
    3. Center-in-Box Constraint: Restricting positive sample centers to lie within the ground-truth bounding box excludes anchors centered on background features outside the object boundaries.
    4. Scale Fairness: In normal-like distributions, approximately 16%16\% of samples lie in [mg+vg,1][m_g + v_g, 1]. Consequently, each ground-truth object receives approximately 0.2×kL0.2 \times kL positive training samples (where LL is the number of feature pyramid levels and kk is candidate count per level), invariant to object scale, aspect ratio, and location. This resolves the scale imbalance of traditional IoU and area-based assignment strategies that assign disproportionately more positive samples to large objects.

  4. Knowl 4 — Redundancy of Multi-Anchor Tiling Under Adaptive Training Sample Selection

    empirical result

    Under traditional IoU-based training sample selection, tiling multiple anchor boxes per location (varying scales and aspect ratios) improves detector performance over tiling a single anchor box per location. On the MS COCO minival dataset, RetinaNet with 9 anchors per location (3 scales×3 aspect ratios3\text{ scales} \times 3\text{ aspect ratios}, denoted #A=9) achieves 38.4%38.4\% AP compared to 37.0%37.0\% AP for 1 anchor per location (#A=1) when incorporating standard modern architectural enhancements.

    However, when Adaptive Training Sample Selection (ATSS) is applied, tiling multiple anchors per location ceases to provide any accuracy benefit:

    Model # Scales # Aspect Ratios AP (%) AP50\text{AP}_{50} (%) AP75\text{AP}_{75} (%)
    RetinaNet (#A=9) Baseline 3 3 36.3 55.2 38.8
    RetinaNet (#A=9) + Improvements 3 3 38.4 56.2 41.6
    RetinaNet (#A=9) + Improvements + ATSS 3 3 39.2 57.6 42.7
    RetinaNet (#A=3) + Improvements + ATSS 3 1 39.3 57.7 42.6
    RetinaNet (#A=3) + Improvements + ATSS 1 3 39.2 57.1 42.5
    RetinaNet (#A=1) + Improvements + ATSS 1 1 39.3 57.5 42.8

    Because adaptive sample selection dynamically selects suitable positive candidates irrespective of the number of preset anchors, tiling multiple anchors per spatial location is redundant.

  5. Knowl 5 — Elimination of Implementation Inconsistencies Between RetinaNet and FCOS

    data/table

    The initial performance gap between anchor-based RetinaNet with one anchor per location (#A=1, 32.5%32.5\% AP) and anchor-free FCOS (37.8%37.8\% AP) on the MS COCO minival dataset is largely driven by universal implementation improvements. Incrementally incorporating these modifications into RetinaNet (#A=1) isolates the core mechanics of each framework:

    Detector GroupNorm GIoU Loss In GT Box Centerness Scalar AP (%)
    FCOS Baseline ✓ ✓ ✓ ✓ ✓ 37.8
    RetinaNet (#A=1) 32.5
    RetinaNet (#A=1) ✓ 33.4
    RetinaNet (#A=1) ✓ ✓ 34.9
    RetinaNet (#A=1) ✓ ✓ ✓ 35.3
    RetinaNet (#A=1) ✓ ✓ ✓ ✓ 36.8
    RetinaNet (#A=1) ✓ ✓ ✓ ✓ ✓ 37.0

    The improvements comprise:

    • GroupNorm: Applying Group Normalization in the classification and regression tower heads (+0.9%+0.9\% AP).
    • GIoU Loss: Replacing Smooth L1 regression loss with Generalized IoU loss (+1.5%+1.5\% AP).
    • In GT Box: Restricting candidate positive samples to be inside the ground-truth box (+0.4%+0.4\% AP).
    • Centerness: Adding an auxiliary branch to down-weight low-quality detections far from object centers (+1.5%+1.5\% AP).
    • Scalar: Introducing a learnable scaling parameter per feature pyramid level for regression targets (+0.2%+0.2\% AP).

    Together, these modifications elevate RetinaNet (#A=1) from 32.5%32.5\% to 37.0%37.0\% AP, reducing the unexplained gap with FCOS to 0.8%0.8\% AP.

  6. Knowl 6 — Performance Impact of ATSS on Anchor-Based and Anchor-Free Architectures

    empirical result

    Applying Adaptive Training Sample Selection (ATSS) to anchor-based (RetinaNet #A=1) and anchor-free (FCOS) detectors on the MS COCO minival dataset improves object detection performance across all metrics without introducing inference overhead:

    Method AP (%) AP50\text{AP}_{50} (%) AP75\text{AP}_{75} (%) APS\text{AP}_S (%) APM\text{AP}_M (%) APL\text{AP}_L (%)
    RetinaNet (#A=1) 37.0 55.1 39.9 21.4 41.2 48.6
    RetinaNet (#A=1) + ATSS 39.3 57.5 42.8 24.3 43.3 51.3
    FCOS Baseline 37.8 55.6 40.7 22.1 41.8 48.8
    FCOS + Center Sampling (Lite ATSS) 38.6 57.4 41.4 22.3 42.5 49.8
    FCOS + ATSS (Full) 39.2 57.3 42.4 22.7 43.1 51.5

    For RetinaNet (#A=1), ATSS yields a +2.3%+2.3\% increase in AP. For FCOS, replacing full ground-truth box spatial candidates with the top k=9k=9 closest points per pyramid level (the lite version, termed center sampling) improves AP by +0.8%+0.8\%. Implementing full ATSS on FCOS (using an 8S8S anchor box at each point to calculate dynamic statistical thresholds while retaining point regression) improves performance by +1.4%+1.4\% AP to 39.2%39.2\%, closing the gap between anchor-based and anchor-free methods.

  7. Knowl 7 — Sensitivity Analysis of ATSS Hyperparameters and Preset Anchor Geometries

    empirical result

    The sensitivity of Adaptive Training Sample Selection (ATSS) with ResNet-50 on MS COCO minival was analyzed across three dimensions:

    1. Candidate Count (kk) per Pyramid Level:
    kk 3 5 7 9 11 13 15 17 19
    AP (%) 38.0 38.8 39.1 39.3 39.1 39.0 39.1 39.2 38.9

    Performance is stable across k∈[7,17]k \in [7, 17] (39.0%−39.3%39.0\% - 39.3\% AP). Low values (k=3k=3, 38.0%38.0\% AP) suffer from statistical instability due to too few candidates, while overly high values (k=19k=19, 38.9%38.9\% AP) introduce low-quality candidate noise.

    1. Preset Anchor Scale (s×Ss \times S, stride SS) with Fixed 1:1 Aspect Ratio:
    • Scale 5S5S: 39.0%39.0\% AP
    • Scale 6S6S: 39.2%39.2\% AP
    • Scale 7S7S: 39.3%39.3\% AP
    • Scale 8S8S: 39.3%39.3\% AP
    • Scale 9S9S: 38.9%38.9\% AP
    1. Preset Anchor Aspect Ratio with Fixed Scale 8S8S:
    • Ratio 4:14:1: 39.1%39.1\% AP
    • Ratio 2:12:1: 39.0%39.0\% AP
    • Ratio 1:11:1: 39.3%39.3\% AP
    • Ratio 1:21:2: 39.3%39.3\% AP
    • Ratio 1:41:4: 39.1%39.1\% AP

    These results demonstrate that ATSS is largely invariant to anchor scale, aspect ratio, and candidate hyperparameter kk within reasonable ranges.

  8. Knowl 8 — Object Detection Performance on MS COCO Benchmark

    data/table

    Performance of ATSS on the MS COCO test-dev benchmark across different backbones and testing configurations, trained using multi-scale jitter (shorter side randomly sampled from 640 to 800) for 180K iterations:

    Method Backbone AP (%) AP50\text{AP}_{50} (%) AP75\text{AP}_{75} (%) APS\text{AP}_S (%) APM\text{AP}_M (%) APL\text{AP}_L (%)
    ATSS ResNet-101 43.6 62.1 47.4 26.1 47.0 53.6
    ATSS ResNeXt-32x8d-101 45.1 63.9 49.1 27.9 48.2 54.6
    ATSS ResNeXt-64x4d-101 45.6 64.6 49.7 28.5 48.9 55.6
    ATSS ResNet-101-DCN 46.3 64.7 50.4 27.7 49.8 58.4
    ATSS ResNeXt-32x8d-101-DCN 47.7 66.6 52.1 29.3 50.8 59.7
    ATSS ResNeXt-64x4d-101-DCN 47.7 66.5 51.9 29.7 50.8 59.4
    ATSS (Multi-scale test) ResNeXt-32x8d-101-DCN 50.6 68.6 56.1 33.6 52.9 62.2
    ATSS (Multi-scale test) ResNeXt-64x4d-101-DCN 50.7 68.9 56.3 33.2 52.9 62.4

    Under single-model and single-scale testing, ATSS with ResNeXt-64x4d-101-DCN reaches 47.7%47.7\% AP. With multi-scale testing, it achieves 50.7%50.7\% AP without introducing additional computation during standard single-scale inference.

Coverage note — No substantial contributed material was omitted; all core methodological insights, experimental comparisons, algorithm details, ablations, and benchmark results are represented.

References

  1. 1.Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. SOD-MTGAN: small object detection via multi-task generative adversarial network. In ECCV, 2018. 8
  2. 2.Sean Bell, C. Lawrence Zitnick, Kavita Bala, and Ross B. Girshick. Inside-outside net: Detecting objects in context with skip pooling and recurrent neural networks. In CVPR, 2016. 2
  3. 3.Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S. Davis. Improving object detection with one line of code. In ICCV, 2017. 8
  4. 4.Zhaowei Cai, Quanfu Fan, Rogerio Schmidt Feris, and Nuno Vasconcelos. A unified multi-scale deep convolutional neural network for fast object detection. In ECCV, 2016. 2
  5. 5.Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN: delving into high quality object detection. In CVPR, 2018. 2, 7, 8
  6. 6.Kean Chen, Jianguo Li, Weiyao Lin, John See, Ji Wang, Lingyu Duan, Zhibo Chen, Changwei He, and Junni Zou. Towards accurate one-stage object detection with ap-loss. In CVPR, 2019. 2
  7. 7.Zhe Chen, Shaoli Huang, and Dacheng Tao. Context refinement for object detection. In ECCV, 2018. 2, 7, 8
  8. 8.Bowen Cheng, Yunchao Wei, Honghui Shi, Rogerio Schmidt Feris, Jinjun Xiong, and Thomas S. Huang. Revisiting RCNN: on awakening the classification power of faster RCNN. In ECCV, 2018. 8
  9. 9.Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-FCN: object detection via region-based fully convolutional networks. In NIPS, 2016. 1, 2, 8
  10. 10.Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 7
  11. 11.Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In ICCV, 2019. 2, 8
  12. 12.Cheng-Yang Fu, Wei Liu, Ananth Ranga, Ambrish Tyagi, and Alexander C. Berg. DSSD : Deconvolutional single shot detector. CoRR, 2017. 2, 8
  13. 13.Ross B. Girshick. Fast R-CNN. In ICCV, 2015. 2
  14. 14.Ross B. Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, 2014. 2
  15. 15.Jiayuan Gu, Han Hu, Liwei Wang, Yichen Wei, and Jifeng Dai. Learning region features for object detection. In ECCV, 2018. 8
  16. 16.Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3
  17. 17.Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, and Xiangyu Zhang. Bounding box regression with uncertainty for accurate object detection. In CVPR, 2019. 2
  18. 18.Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, 2018. 8
  19. 19.Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, and Kevin Murphy. Speed/accuracy trade-offs for modern convolutional object detectors. In CVPR, 2017. 8
  20. 20.Lichao Huang, Yi Yang, Yafeng Deng, and Yinan Yu. Densebox: Unifying landmark localization with end to end object detection. CoRR, 2015. 3
  21. 21.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, 8
  22. 22.Tao Kong, Fuchun Sun, Wen-bing Huang, and Huaping Liu. Deep feature pyramid reconfiguration for object detection. In ECCV, 2018. 2
  23. 23.Tao Kong, Fuchun Sun, Huaping Liu, Yuning Jiang, and Jianbo Shi. Foveabox: Beyond anchor-based object detector. CoRR, 2019. 1, 3, 8
  24. 24.Tao Kong, Fuchun Sun, Anbang Yao, Huaping Liu, Ming Lu, and Yurong Chen. RON: reverse connection with objectness prior networks for object detection. In CVPR, 2017. 2
  25. 25.Tao Kong, Anbang Yao, Yurong Chen, and Fuchun Sun. Hypernet: Towards accurate region proposal generation and joint object detection. In CVPR, 2016. 2
  26. 26.Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, 2018. 1, 2, 8
  27. 27.Hei Law, Yun Teng, Olga Russakovsky, and Jia Deng. Cornernet-lite: Efficient keypoint based object detection. CoRR, 2019. 2, 8
  28. 28.Hyungtae Lee, Sungmin Eum, and Heesung Kwon. ME R-CNN: multi-expert region-based CNN for object detection. In ICCV, 2017. 2
  29. 29.Shuai Li, Lingxiao Yang, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Dynamic anchor feature selection for single-shot object detection. In ICCV, 2019. 2
  30. 30.Yanghao Li, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Scale-aware trident networks for object detection. In ICCV, 2019. 2
  31. 31.Zeming Li, Chao Peng, Gang Yu, Xiangyu Zhang, Yangdong Deng, and Jian Sun. Detnet: A backbone network for object detection. In ECCV, 2018. 8
  32. 32.Tsung-Yi Lin, Piotr Dollar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 1, 2
  33. 33.Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 1, 2, 3, 7, 8
  34. 34.Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 2, 3
  35. 35.Songtao Liu, Di Huang, and Yunhong Wang. Receptive field block net for accurate and fast object detection. In ECCV, 2018. 2, 8
  36. 36.Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott E. Reed, Cheng-Yang Fu, and Alexander C. Berg. SSD: single shot multibox detector. In ECCV, 2016. 1, 2, 8
  37. 37.Wei Liu, Shengcai Liao, Weiqiang Ren, Weidong Hu, and Yinan Yu. High-level semantic feature detection: A new perspective for pedestrian detection. In CVPR, 2019. 3
  38. 38.Yong Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. Structure inference net: Object detection using scene-level context and instance-level relationships. In CVPR, 2018. 2
  39. 39.Xin Lu, Buyu Li, Yuxin Yue, Quanquan Li, and Junjie Yan. Grid R-CNN. In CVPR, 2019. 2, 8
  40. 40.Mahyar Najibi, Mohammad Rastegari, and Larry S. Davis. G-CNN: an iterative grid based object detector. In CVPR, 2016. 2
  41. 41.Mahyar Najibi, Bharat Singh, and Larry S. Davis. Autofocus: Efficient multi-scale inference. In ICCV, 2019. 2
  42. 42.Jing Nie, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Enriched feature guided refinement network for object detection. In ICCV, 2019. 2
  43. 43.Jiangmiao Pang, Kai Chen, Jianping Shi, Huajun Feng, Wanli Ouyang, and Dahua Lin. Libra R-CNN: towards balanced learning for object detection. In CVPR, 2019. 2
  44. 44.Zheng Qin, Zeming Li, Zhaoning Zhang, Yiping Bao, Gang Yu, Yuxing Peng, and Jian Sun. Thundernet: Towards real-time generic object detection. In ICCV, 2019. 2
  45. 45.Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 2
  46. 46.Joseph Redmon and Ali Farhadi. YOLO9000: better, faster, stronger. In CVPR, 2017. 8
  47. 47.Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. TPAMI, 2017. 1, 2
  48. 48.Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian D. Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 3
  49. 49.Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Fei-Fei Li. Imagenet large scale visual recognition challenge. IJCV, 2015. 3
  50. 50.Zhiqiang Shen, Zhuang Liu, Jianguo Li, Yu-Gang Jiang, Yurong Chen, and Xiangyang Xue. DSOD: learning deeply supervised object detectors from scratch. In ICCV, 2017. 2
  51. 51.Abhinav Shrivastava and Abhinav Gupta. Contextual priming and feedback for faster R-CNN. In ECCV, 2016. 2
  52. 52.Abhinav Shrivastava, Abhinav Gupta, and Ross B. Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016. 2
  53. 53.Abhinav Shrivastava, Rahul Sukthankar, Jitendra Malik, and Abhinav Gupta. Beyond skip connections: Top-down modulation for object detection. CoRR, 2016. 8
  54. 54.Bharat Singh and Larry S. Davis. An analysis of scale invariance in object detection - SNIP. In CVPR, 2018. 2, 7, 8
  55. 55.Zhiyu Tan, Xuecheng Nie, Qi Qian, Nan Li, and Hao Li. Learning to rank proposals for object detection. In ICCV, 2019. 2
  56. 56.Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: fully convolutional one-stage object detection. In ICCV, 2019. 1, 2, 3, 7, 8
  57. 57.Lachlan Tychsen-Smith and Lars Petersson. Improving object localization with fitness NMS and bounded iou loss. In CVPR, 2018. 8
  58. 58.Hao Wang, Qilong Wang, Mingqi Gao, Peihua Li, and Wangmeng Zuo. Multi-scale location-aware kernel representation for object detection. In CVPR, 2018. 8
  59. 59.Jiaqi Wang, Kai Chen, Shuo Yang, Chen Change Loy, and Dahua Lin. Region proposal by guided anchoring. In CVPR, 2019. 3, 8
  60. 60.Tiancai Wang, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Learning rich features at high-speed for single-shot object detection. In ICCV, 2019. 2
  61. 61.Xiaolong Wang, Abhinav Shrivastava, and Abhinav Gupta. A-fast-rcnn: Hard positive generation via adversary for object detection. In CVPR, 2017. 2
  62. 62.Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 3
  63. 63.Saining Xie, Ross B. Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 7
  64. 64.Hongyu Xu, Xutao Lv, Xiaoyu Wang, Zhou Ren, Navaneeth Bodla, and Rama Chellappa. Deep regionlets for object detection. In ECCV, 2018. 8
  65. 65.Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In ICCV, 2019. 2, 8
  66. 66.Shifeng Zhang, Longyin Wen, Xiao Bian, Zhen Lei, and Stan Z. Li. Single-shot refinement neural network for object detection. In CVPR, 2018. 2, 7, 8
  67. 67.Xiaosong Zhang, Fang Wan, Chang Liu, Rongrong Ji, and Qixiang Ye. Freeanchor: Learning to match anchors for visual object detection. In NIPS, 2019. 2
  68. 68.Zhishuai Zhang, Siyuan Qiao, Cihang Xie, Wei Shen, Bo Wang, and Alan L. Yuille. Single-shot object detection with enriched semantics. In CVPR, 2018. 2, 8
  69. 69.Peng Zhou, Bingbing Ni, Cong Geng, Jianguo Hu, and Yi Xu. Scale-transferrable object detection. In CVPR, 2018. 2, 8
  70. 70.Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. CoRR, 2019. 2, 8
  71. 71.Xingyi Zhou, Jiacheng Zhuo, and Philipp Krahenbuhl. Bottom-up object detection by grouping extreme and center points. In CVPR, 2019. 1, 2, 8
  72. 72.Chenchen Zhu, Yihui He, and Marios Savvides. Feature selective anchor-free module for single-shot object detection. In CVPR, 2019. 3, 8
  73. 73.Rui Zhu, Shifeng Zhang, Xiaobo Wang, Longyin Wen, Hailin Shi, Liefeng Bo, and Tao Mei. Scratchdet: Exploring to train single-shot object detectors from scratch. CoRR, 2018. 2
  74. 74.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. 8

Citation

MLA
Zhang, S., et al. “Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection”. arXiv, 2019, http://arxiv.org/abs/1912.02424v4.
APA
Zhang, S., Chi, C., Yao, Y., Lei, Z., & Li, S. Z. (2019). Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection. arXiv. http://arxiv.org/abs/1912.02424v4
Chicago
Zhang, S., C. Chi, Y. Yao, Z. Lei, and S. Z. Li. 2019. “Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection”. arXiv. http://arxiv.org/abs/1912.02424v4.
Harvard
Zhang, S. et al. (2019) “Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1912.02424v4.
Vancouver
1. Zhang S, Chi C, Yao Y, Lei Z, Li SZ (2019) Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection. arXiv

BibTeX

@article{zhang2019bridging,
  title = {Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection},
  author = {Zhang, Shifeng and Chi, Cheng and Yao, Yongqiang and Lei, Zhen and Li, Stan Z.},
  year = {2019},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1912.02424v4},
  eprint = {1912.02424}
}
Metadata:arXiv

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: IEEE