框架:SegNet, a deep convolutional encoder-decoder architecture
框架组成部分:A stack of encoders
+ a corresponding decoder stack
+ then feed into a softmax classification layer
应用:semantic pixel wise image labeling
三个亮点:1、only require forward evaluation of a fully learnt function to obtain smooth label predictions
(仅需要一个前向估计函数就可以获得平滑标签预测)
2、with increasing depth (of neural network), a larger context is considered for pixel labeling which improves accuracy.
(网络深度增加,像素点context信息能够考虑到)
3、be easy to visualizes the effect of feature activations in the pixel labels space at any depth.
(在网络的任何一层,都可以可视化特征)
Framework