摘要：
本文提出了一个速度和精度权衡的detector。主要研究两个问题？
how to make the anchor-free detection head better?
how to utilize the power of feature pyramid better?
分别提出两种训练策略：soften optimization techniques，
1）soft-weighted anchor points
2）soft-selected pyramid levels
实验表明，我们简洁的SAPD将速度/精度权衡提升到了一个新的水平，COCO上获得a single-model single-scale AP of 47.4%，5xfaster than other detectors.
———————————————————————————————————————————————
目标检测分为anchor-base和anchor-free，anchor-free主要分为两大类，anchor-point和key-point检测。
Anchor-point方法：
将bounding box根据point-to-boundary distance编解码成anchor points 这些anchor point=pixel on pyramid feature maps并具有相应的距离信息。
优点：1.框架简单 2.训练速度快 3.backbone增大有帮助 4. flexible feature level selection
缺点：same image scale ->low mAP
Key-point方法：
预测bounding box的关键点。
优点：input image size small-> high mAP
缺点：下采样不能过多，需要single high-resolution and repeat bottom-down ，down-bottom in processing.因此FLOPs大，训练时间长，计算内存消耗大，对预先练backbone兼容性差。
没有DCN（deformable convolution）5x faster
有DCN（deformable convolution）high mAP
一般的Anchor-point detector模型框架
1.Feature Pyramid 中l代表input-size*(1/sl),sl=2**l
2.Head中有两个子网络，分类（每个anchor-point输出K类的可能性）和回归（4个值)
3.Supervision targets：Plij表示在第l层，（i，j）坐标的特征， i = 0,1,…,W/sl − 1 j = 0,1,…,H/sl −1. Each plij has a corresponding image space location (Xlij,Ylij) where Xlij = sl(i + 0.5) and Ylij = sl(j + 0.5)。
正样本判断：ground-truth B = (c,x,y,w,h)，中心收缩Bv = (c,x,y,ew,eh), 只有在BV里面才是positive。
最后归一化的left，top，right，behind的坐标为：
- Loss functions ：focal-loss 分类，IOU-loss 回归
两个创新策略
1.SW（soft-weight ）
权重公式：yita控制下降步伐。
2.ss（soft-select）
Instance-dependent经过每层level->ROI Align->concat->经过meta-selection network-> a vector
The meta-selection network is jointly trained with the detector. Cross entropy loss is used for optimization and the ground-truth is a one hot vector indicating which pyramid level has minimal loss as defined in the FSAF module
Loss
———————————————————————————————————————————————
消融实验
结果表明，对于金字塔级别更高的实例，更大的实例往往被赋予更高的权重。大多数事例不超过两层就可以学会。非常罕见的实例需要超过两层，例如图7右上角的沙发子图。这与表4中的结果一致。
Joint training of the meta-selection network has a negligible effect on performance.
有人认为是因为多任务训练产生优异的表现，但是本文只是用训练进行联合训练，验证并没有使用权重，换句话说特征选择策略 Is the same as the baseline FSAF module
SAPD is robust and efficient
修改backbone很容易提高AP，AR
并且和anchor-base方法进行比较，实验表明
SAPD不仅跑的快（因为head 结构简单）
并且精准率高（即使是anchor-base和anchor-free方法结合
by significant margins）
Without DCN, our fastest SAPD version based on ResNet-50 can reach a 14.9 FPS while maintaining a 41.7% AP
With DCN, our SAPD forms an upper envelope of recent state-of-the-art anchor-based and anchor-free detectors
———————————————————————————————————————————————
补充
FSAF