摘要
道路目标检测是自动驾驶环境感知的重要组成部分。现实场景中数据往往是遵循长尾分布的,数据的长尾分布会导致分类器过度拟合样本数量多的类别,使得算法对于样本数量少的类别检测效果较差。为了缓解这一问题提出了SFL(seesaw focal loss),使用互补的缓解因子与补偿因子动态地平衡每个类别的正样本和负样本梯度。缓解因子减少对尾部类别的惩罚,补偿因子增加了对错误分类的惩罚,以避免尾部类别的误检。基于SFL和EfficientDet提出了SEfficientDet系列目标检测算法。通过在BDD100K数据集上的实验表明,SEfficientDet可以显著提升尾部类别的识别效果。与EfficientDet相比,SEfficientDet在同等参数量下mAP提高4.8%,训练速度提高2.3倍。在同等推理时间下mAP提高5.3%,训练速度提升1.8倍。
Road object detection is an important part of automatic driving environment perception.In real scenarios,data often follows a long-tailed distribution.The long-tailed distribution of data will cause the classifier to overfit categories with a large number of samples,making the algorithm’s detection effect for tail categories poor.To alleviate this problem,this paper proposes SFL(seesaw focal loss),which uses complementary mitigation factors and compensation factors to dynamically balance the positive and negative sample gradients of each category.The mitigation factor reduces punishments to tail categories,the compensation factor increases the penalty of misclassified instances to avoid false positives of tail categories.Based on SFL and EfficientDet,this paper proposes the SEfficientDet series of object detection algorithms.Experiments on the BDD100Kdataset show that SEfficientDet can significantly improve the recognition effect of tail categories.Compared with EfficientDet,SEfficientDet improves mAP by 4.8%and training speed by 2.3 times under the similar parameters.Under the similar inference time,mAP is increased by 5.3%,and the training speed is increased by 1.8 times.
作者
王志红
王煜晟
WANG Zhi-hong;WANG Yu-sheng(Hubei Key Laboratory of Advanced Technology for Automotive Components(Wuhan University of Technology),Wuhan 430070,China;Hubei Collaborative Innovation Center for Auto Parts Technology,Wuhan 430070,China)
出处
《武汉理工大学学报》
CAS
2022年第10期102-108,共7页
Journal of Wuhan University of Technology
基金
新能源汽车科学与关键技术学科创新引智基地(B17034)
教育部创新团队发展计划(IRT_17R83)
作者简介
王志红(1980-),男,讲师,硕导.E-mail:wangzhihong@whut.edu.cn
