Conditional Random Fields as Recurrent Neural Networks_2015

Shuai Zheng , Sadeep Jayasumana , Bernardino RomeraParedes - IEEE International Conference on Computer Vision - 2015 - 被引量:447

目录结构：

摘要

1.介绍

2.相关工作

3.条件随机场

4.A Mean-field Iteration as a Stack of CNN Layers

4.1初始化

4.2消息传递（Message Passing） 

4.3滤波器输出（Weighting Filter Outputs）

4.4兼容性转换（Compatibility Transform）

4.5加上一元势函数（Adding Unary Potentials）

4.6正则化（Normalization）

5.端对端的可训练的网络

5.1CRF as RNN

5.2完善图像（Completing the Picture）

6.实现细节

7.实验

Pascal VOC Datasets
Pascal Context Dataset

7.1. Effect of Design Choices

8.结论

摘要
1.介绍
2.相关工作
3.条件随机场
4.A Mean-field Iteration as a Stack of CNN Layers

A key contribution of this paper is to show that the mean- field CRF inference can be reformulated as a RNN.

这篇论文的一篇的主要贡献就是把mean-field CRF的推理过程用RNN来阐述。

To this end, we first consider individual steps of the mean-field algorithm summarized in Algorithm 1 [29], and describe them as CNN layers.

我们第一次把mean-field算法总结到算法1中。

迭代终止的条件是，谁要收敛？

Our contribution is based on the observation that filter-based approximate mean-field inference approach for dense CRFs relies on applying Gaussian spatial and bilateral filters on the mean-field approximates in each iteration. 

我们的主要贡献是基于这样的观察，filter-based approximate mean-field inference approach for dense CRFs依赖于将高斯空间和双边滤波器应用于每次迭代上的mean-field approximates。

Unlike the standard convolutional layer in a CNN, in which filters are fixed after the training stage, we use edge-preserving Gaussian filters [56, 42], coefficients of which depend on the original spatial and appearance information of the image.

不像CNN中标准的卷积层，这儿的滤波器都是在训练后被放置，我们使用保留边缘的高斯滤波器[56,42]。

These filters have the additional advantages of requiring a smaller set of parameters, despite the filter size being potentially as big as the image.

这些滤波器有着额外的优势，即需要一组更少的参数，尽管滤波器的大小可能和图像一样大。

4.1初始化
4.2消息传递（Message Passing） 
4.3滤波器输出（Weighting Filter Outputs）
4.4兼容性转换（Compatibility Transform）
4.5加上一元势函数（Adding Unary Potentials）
4.6正则化（Normalization）
5.端对端的可训练的网络
5.1CRF as RNN
5.2完善图像（Completing the Picture）
6.实现细节
7.实验
Pascal VOC Datasets
Pascal Context Dataset
7.1. Effect of Design Choices
8.结论