Hei Law, Yun Teng, Olga Russakovsky, Jia Deng
Princeton University, 2019

### Detecting Objects

The locations obtained from the bounding box predictions give more information about the object sizes. We can use the sizes of the bounding boxes to determine zoom-in scales. The scale is determined such that the longer side of the bounding box after zoom-in is 24 for a small object, 64 for a medium object and 192 for a large object.

### Backbone Network

Hourglass-54里的每个module都比104里浅、参数少。下采样的步长为2。每个Hourglass模块会下采样特征图3次，并按(384,384,512)来增加通道数。模块中有一个512通道的残差模块。

## CornerNet-Squeeze

SqueezeNet共使用了3中策略来降低网络复杂度：(1) 使用$1\times1$替换$3\times3$ (2) decreasing input channels to $3\times3$ kernels (3) 晚一点下采样。