ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation
(a)是intial block。Input为512x512。Num_filters为13，max_pool为2x2，s=2，concat变成16x512x512
(b)bottleneck module：1x1conv reduce dim + conv + 1x1conv expend dim。Bn+PRelu between all convs。
If bottleneck module在downsample，需要在mian branch 添加 maxpooling，同时，第一个1x1conv replace为2*2conv，s=2（相当于进行pool缩小一半），
Conv可用正常conv，dilate conv，全卷积，3x3。有时用5x1 and 1x5conv
对于regularizer正则化，dropout在bottleneck之前p=0.01，之后p=0.1。使用了segnet的方法，save pool的索引，然后upsample。
(1)downsampling will reduce feature map resolution, it implies loss of spatial information like exact edge shape. Filters operating with downsampled images have a bigger receptive field , that allows them to gather more context.(dilated conv)
(2)Early downsample，减少input size。他认为初始是网络层并不直接作用于分类，只是充当好的特征提取器
(3)Decoder size。本文较大的encoder较小的decoder。他认为encoder是对数据的信息进行处理，而decoder只是将encoder上采样，只是调整图像的细节
