Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Segmentation #5

Open
Kinpzz opened this issue Jul 24, 2017 · 4 comments
Open

Segmentation #5

Kinpzz opened this issue Jul 24, 2017 · 4 comments

Comments

@Kinpzz
Copy link
Owner

Kinpzz commented Jul 24, 2017

Deeplab 图像分割方法

参考博客:

@Kinpzz Kinpzz changed the title deeplab Segmentation Aug 9, 2017
@Kinpzz
Copy link
Owner Author

Kinpzz commented Aug 9, 2017

FCN与CNN的区别

FCN的最后几层不是全连接层,而CNN是全连接层。这些全连接层都是一维的信息,丢失了二维的信息。
CNN是图像到结果的网络,从二维信息到一维信息,一般输出一个图片的结果。而FCN网络是从图像到图像的网络,从二维信息到二维信息,是一个像素级的网络,对应每个像素点的结果。

CNN的识别是图像级的识别,也就是从图像到结果,而FCN的识别是像素级的识别,对输入图像的每一个像素在输出上都有对应的判断标注,标明这个像素最可能是属于一个什么物体/类别。

上采样和下采用

CNN从高维度到低维度,卷积+池化的过程就是下采用。FCN从低维度到高维度,反卷积的过程就是上采样。

bilinear interpolation 双线性内插

感受野

这里就涉及到一个感受野(receptive field)的概念。较浅的卷积层(靠前的)的感受域比较小,学习感知细节部分的能力强,较深的隐藏层(靠后的),感受域相对较大,适合学习较为整体的,相对更宏观一些的特征。
感受野被定义为特定 CNN 特征正在“看”(即受其影响)的输入空间中的区域。特征的感受野可以通过其中心位置及其大小进行充分描述。

所以在做反卷积的时候,会考虑浅层的卷积信息,辅助叠加得到更好的分割结果。

概念

  • 条件随机场
  • Dilated Convolutions

@Kinpzz
Copy link
Owner Author

Kinpzz commented Aug 10, 2017

FCN

Introduction

Title: Fully Convolutional Networks for Semantic Segmentation
Author: Jonathan Long, Evan Shelhamer, Trevor Darrell, UC Berkeley
arXiv: 1411.4038
2015 CVPR Best Paper Honorable Mention

#Keyword

  • semantic segmentation(语义分割)
  • end-to-end, pixels-to-pixels(端到端,像素级)
  • fully convolutional networks

(1) 将最后几层全连接层改为卷积层,大小为原来1d全连接层进行resize(故称“全卷积网络”)
(2) 将入了转置卷积,并与浅层池化层输出结合,通过设置stride来放大

Main work

  • make predition a every pixel
  1. 全连接层解释为卷积层,fine-tune
  2. upsampleing(eg. bilinear upsampling) (TODO)
    内插法,上采样(多的位置填0)

@Kinpzz
Copy link
Owner Author

Kinpzz commented Aug 11, 2017

Deeplab

@Kinpzz
Copy link
Owner Author

Kinpzz commented Aug 11, 2017

SegNet

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant