人群分割--Fully Convolutional Neural Networks for Crowd Segmentation

Fully Convolutional Neural Networks for Crowd Segmentation
https://arxiv.org/abs/1411.4464

这里设计了一个全卷积网络用于视频中的人群分割，主要考虑三个信息：Apperance、 Motion 、Structure，思路还是很原始的。
主要的难度在于 静态的人群我们也想分割出来，再就是当人群的纹理和背景相似的时候，这个时候就需要靠运动信息了。

CNN and FCNN models

Multi-stage fusion structure

三个网络结构一样的：The our proposed FCNN can be represented as: Conv(32,7,1) - ReLU - Pool(MAX,2,2) -Conv(64,7,1) - ReLU - Pool(MAX,2,2) - Conv(128,3,1) -ReLU - Conv(128,3,1) - ReLU - Conv(64,3,1) - ReLU -Conv(16,3,1) - ReLU - Conv(1,1,1) - Sig.

Fusion Schemes and Extensible Architecture
input fusion, feature fusion, and decision fusion.

1） The input fusion directly concatenate input maps as multiple channels.
2） The feature fusion combines output feature maps of a certain fusion layer and use feature maps of all three networks to make a decision.
3）The decision fusion scheme combines the output maps of three separately trained networks.

训练数据标记

多视角训练