1. Motivation
anchor-based method(RetinaNet)和 anchor-free method(FCOS)的主要差异体现在以下4点:
-
The number of anchors tiled per location.
-
The definition of positive and negative samples.
-
The regression starting status.
而目前FCOS的实验结果会比RetinaNet好,因此在这三个差异中,哪一点是造成这2个detectors的性能gap的主要因素值得研究。
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 be- tween them.
在本文中,作者调查了anchor-based和anchor-free2种目标检测方法在精度上的的差异。通过排除2种方法在实现上不一致的地方,可以得出结论,2种方法在性能上的gap来源于正负样本的定义。
如何更好的制定正负样本值得进一步的研究,基于此动机,本文提出了自适应训练样本选择器(ATSS),基于物体的特征来自动的选择正负样本。
Then, we propose an Adaptive Training Sample Selection (ATSS) to automatically select positive and negative samples according to statistical characteristics of object.
2. Contribution
本文的主要贡献可以可以总结为以下4点:
- 实验表明,anchor-based和anchor-free detectors的差异性主要是源于正负样本的制定上。
- 提出了ATSS方法,根据物体的statistical characteristic