FeatureNMS-- the rule whether to add p from P to D or not is adjusted

FeatureNMS
算法流程图

具体改进

In any other case the two bounding boxes might belong to the same or to different objects—the intersection over union alone cannot be used to make a final decision. In this case we calculate the L2 distance of feature embeddings for both bounding boxes. If this distance is larger than a threshold T we assume that the bounding boxes belong to different objects. Otherwise they are likely to belong to the same object

用两个阈值来控制上述方框内的内容，使得对于IOU小于N2的还能根据嵌入特征进行二次筛选，如果比较p和d的嵌入特征L2距离大于一个阈值，说明他与p不是同一个类

: The proposal p with the highest confidence score in P

嵌入特征提取:

We add one head to the RetinaNet backbone. This head outputs a feature
embedding for each anchor box. We chose an embedding of length of 32,
but other lengths are possible.