面向小目标检测的扩展特征金字塔网络
计算机视觉与模式识别
2020-04-10 v2
摘要
小目标检测仍是一项未解难题,因为仅含极少像素的小目标难以提取信息。尽管特征金字塔网络中的尺度级对应检测缓解了此问题,我们发现不同尺度的特征耦合仍会损害小目标性能。本文提出扩展特征金字塔网络(EFPN),其带有专为小目标检测设计的额外高分辨率金字塔层级。具体而言,我们设计了一个名为特征纹理迁移(FTT)的新模块,用于同时超分辨率化特征并提取可信区域细节。此外,我们设计了前景-背景平衡损失函数以缓解前景与背景的面积失衡。实验中,所提 EFPN 在计算与内存上均高效,并在小交通标志数据集 Tsinghua-Tencent 100K 与通用目标检测数据集 MS COCO 的小目标类别上取得最先进结果。
引用
@article{arxiv.2003.07021,
title = {Extended Feature Pyramid Network for Small Object Detection},
author = {Chunfang Deng and Mengmeng Wang and Liang Liu and Yong Liu},
journal= {arXiv preprint arXiv:2003.07021},
year = {2020}
}
备注
With the agreement of all authors, we would like to withdraw the manuscript. For lack of some experiments, a part of important claims cannot stand solidly. We need to further carry out experiments, and reconsider the rationality of these claims