1、features in deeper convolution layer output have wider reception fields
2、迭代次数=（图片数目/batch_size）*epoch
3、Here is the dilemma – On one hand, more convolution layers are required to learn highly representative features that can distinguish faces with large appearance variations, i.e., pose, expression, and occlusion from clutter background. On the other hand, by going deeper, the spatial information will lose through pooling or convolution operations. This spatial information is essential to recognize tiny objects.
-----Face Detection through Scale-Friendly Deep Convolutional Networks
4、ScaleFace can be compressed by reducing the number of filters in every layer of the backbone network.
-----scaleface
5、network structure and target scale are correlated-----scaleface
6、较浅的卷积层感知域较小，学习到一些局部区域的特征；较深的卷积层具有较大的感知域，能够学习到更加抽象一些的特征。这些抽象特征对物体的大小、位置和方向等敏感性更低，从而有助于识别性能的提高。