cv-segNet学习笔记

显著特点

segNet的特点最显著特点就是decoder网络复用了对应的encoder网络的降采样层来升采样（即max-pooling indices）

文章结构

section2 最近成就
section3 segNet及分析
seciton4 评价segNet表现
section5 未来工作
section6 结论

section2

before deep network: hand engineering features + classifier

CamVid

1. SfM+Random Forest/Boosting
2. CRF

Indoor NYUdataset

1. RGB-SIFT/depth-SIFT/piexel location/LBP
2. resl-time joint reconstruction and semantic segmentation RandomForest

deep CNN

1. 复制分类特征用于分割replicating the deepest layer features in blocks to match image
2. merge rnn 低分辨率
3. encoder+decoder network
4. CRF-RNN
5. 多规格深度结构:a.使用少量规模和对应featuremap;b,结合不同层的featuremap

section3

1. encoder + 对应的decoder