一种基于锚框质量分布的动态标签分配策略

doi:10.19678/j.issn.1000-3428.0064346

摘要/Abstract

摘要： 传统的标签分配策略根据锚框与ground-truth之间的交并比（IoU）是否超过阈值来划分正负样本，但确定IoU阈值需要依靠大量的经验和实验，并且在不同的网络结构中需要重新设定新的阈值；其次固定的阈值无法保证正负样本数量变化的平缓，过多或过少的正样本都将影响网络训练的稳定性。为解决这些问题，提出一种自适应标签分配策略。根据中心先验原则设计中心权重，通过联合分类得分和定位得分表示锚框质量，组成一个统一的锚框评分方案便于简化划分步骤。为了使网络能够根据训练程度自适应调整IoU阈值，保证网络训练的稳定性，利用锚框质量高斯函数模拟锚框总体质量得分的分布情况，使用极大似然估计算法获取最优参数，根据统计结果动态预测最佳IoU阈值。实验结果表明，该算法在基于锚框和基于无锚框的检测算法中均使检测性能得到有效提升，在PASCAL VOC数据集上分别提升3.1和6.6个百分点，并且可以有效降低漏检率。

关键词: 标签分配, 目标检测, 深度学习, 样本, 交并比

Abstract: Traditional label assignment strategies divide positive and negative samples according to whether the Intersection over Union(IoU) between the anchor and ground-truth exceeds the threshold.However, determining the IoU threshold requires considerable experience and experimentation, and new thresholds must be set for different network structures.Moreover, a fixed threshold cannot guarantee the consistent partitioning of positive and negative samples, and too many or too few positive samples will affect the reliability of network training.To solve these problems, we propose an Adaptive Label Assignment Strategy(ALAS).First, the center weights are designed according to the center prior principle;the anchor quality is represented by the joint classification and localization scores;and a unified anchor scoring scheme is developed to simplify the partition.To enable the network to adaptively adjust the IoU threshold according to the degree of training and ensure the consistency of network training, the anchor quality Gaussian function is used to simulate the distribution of the overall quality score of the anchor.The maximum likelihood estimation algorithm is used to obtain optimal parameters, and then the optimal IoU threshold is dynamically predicted based on statistical results.The experimental results show that the algorithm can effectively improve both anchor-based and non-anchor-based detection algorithms, and it improves these algorithms by 3.1 and 6.6 percentage points, respectively, on the PASCAL VOC dataset. Therefore, it can effectively reduce missed detection rates.

Key words: label assignment, object detection, deep learning, sample, Intersection over Union(IoU)

中图分类号:

TP18

王璐璐, 陈东方, 王晓峰. 一种基于锚框质量分布的动态标签分配策略[J]. 计算机工程, 2023, 49(4): 85-91,100.

WANG Lulu, CHEN Dongfang, WANG Xiaofeng. A Dynamic Label Assignment Strategy Based on Quality Distribution of Anchor[J]. Computer Engineering, 2023, 49(4): 85-91,100.

https://www.ecice06.com/CN/Y2023/V49/I4/85

图/表 10

20230415183449

20230415183452

20230415183455

20230415183458

20230415183501

20230415183505

20230415183508

20230415183511

20230415183514

20230415183518

参考文献

[1] 张晨阳, 黄腾, 吴壮壮.基于K-Means聚类与深度学习的RGB-D SLAM算法[J].计算机工程, 2022, 48(1):236-244, 252. ZHANG C Y, HUANG T, WU Z Z.RGB-D SLAM algorithm based on K-means clustering and deep learning[J].Computer Engineering, 2022, 48(1):236-244, 252.(in Chinese)
[2] JIANG B, LUO R, MAO J, et al.Acquisition of localization confidence for accurate object detection[C]//Proceedings of European Conference on Computer Vision.New York, USA:ACM Press, 2018:816-832.
[3] 包俊, 刘宏哲.融合可变形卷积网络的鱼眼图像目标检测[J].计算机工程, 2021, 47(4):248-255. BAO J, LIU H Z.Object detection in fisheye images combining deformable convolutional networks[J].Computer Engineering, 2021, 47(4):248-255.(in Chinese)
[4] LIN T Y, GOYAL P, GIRSHICK R, et al.Focal loss for dense object detection[C]//Proceedings of IEEE International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2017:2999-3007.
[5] ZHANG X S, WAN F, LIU C, et al.FreeAnchor:learning to match anchors for visual object detection[EB/OL].[2022-02-10].https://arxiv.org/abs/1909.02466.
[6] TIAN Z, SHEN C H, CHEN H, et al.FCOS:fully convolutional one-stage object detection[C]//Proceedings of IEEE/CVF International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2020:9626-9635.
[7] ZHANG S F, CHI C, YAO Y Q, et al.Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2020:9756-9765.
[8] ZHU B J, WANG J F, JIANG Z K, et al.AutoAssign:differentiable label assignment for dense object detection[EB/OL].[2022-02-10].https://arxiv.org/abs/2007. 03496.
[9] ZHU C C, HE Y H, SAVVIDES M.Feature selective anchor-free module for single-shot object detection[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2020:840-849.
[10] REDMON J, FARHADI A.YOLO9000:better, faster, stronger[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2017:6517-6525.
[11] KIM K, LEE H S.Probabilistic anchor assignment with IoU prediction for object detection[C]//Proceedings of European Conference on Computer Vision.Berlin, Germany:Springer, 2020:355-371.
[12] LIN T Y, MAIRE M, BELONGIE S, et al.Microsoft COCO:common objects in context[C]//Proceedings of European Conference on Computer Vision.Berlin, Germany:Springer, 2014:740-755.
[13] BÜHLMANN P, YU B.Boosting with the L2 Loss[J].Journal of the American Statistical Association, 2003, 98(462):324-339.
[14] REDMON J, DIVVALA S, GIRSHICK R, et al.You only look once:unified, real-time object detection[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2016:779-788.
[15] REN S Q, HE K M, GIRSHICK R, et al.Faster R-CNN:towards real-time object detection with region proposal networks[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6):1137-1149.
[16] EVERINGHAM M, VAN GOOL L, WILLIAMS C K I, et al.The pascal Visual Object Classes (VOC) challenge[J].International Journal of Computer Vision, 2010, 88(2):303-338.
[17] HE K M, ZHANG X Y, REN S Q, et al.Deep residual learning for image recognition[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2016:770-778.
[18] DAI J F, QI H Z, XIONG Y W, et al.Deformable convolutional networks[C]//Proceedings of IEEE International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2017:764-773.
[19] LI Z M, PENG C, YU G, et al.DetNet:design backbone for object detection[C]//Proceedings of European Conference on Computer Vision.Berlin, Germany:Springer, 2018:339-354.
[20] BODLA N, SINGH B, CHELLAPPA R, et al.Soft-NMS-improving object detection with one line of code[C]//Proceedings of IEEE International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2017:5562-5570.
[21] LI H D, WU Z X, ZHU C, et al.Learning from noisy anchors for one-stage object detection[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2020:10585-10594.
[22] ZHU C C, HE Y H, SAVVIDES M.Feature selective anchor-free module for single-shot object detection[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2020:840-849.
[23] ZHU C C, CHEN F Y, SHEN Z Q, et al.Soft anchor-point object detection[C]//Proceedings of European Conference on Computer Vision.Berlin, Germany:Springer, 2020:91-107.
[24] KE W, ZHANG T L, HUANG Z Y, et al.Multiple anchor learning for visual object detection[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington D.C., USA:IEEE Press, 2020:10203-10212.
[25] YANG Z, LIU S H, HU H, et al.RepPoints:point set representation for object detection[C]//Proceedings of IEEE/CVF International Conference on Computer Vision.Washington D.C., USA:IEEE Press, 2020:9656-9665.
[26] LIU W, ANGUELOV D, ERHAN D, et al.SSD:single shot multibox detector[C]//Proceedings of European Conference on Computer Vision.Berlin, Germany:Springer, 2016:21-37.

选择文件类型/文献管理软件名称

选择包含的内容