论文思路
论文：<<Bayesian color constancy >>
Title&&Abstract
MSE:
https://blog.csdn.net/qq_36512295/article/details/86526799
MMSE:
the Maximum Local Mass Estimate
什么是局部最大质量估计？
Figure
Figure1
We assume that each surface is flat and matte, so that it may be characterized by a single spectral reflectance function
为什么要flat and matte?
平坦是保证处处一致，无光泽不光滑是保证是漫反射吧应该？
The spectral power distri- bution of the light reaching the observer from each surface is given as the wavelength-by-wavelength product of the illuminant spectral power distribution and the surface reflectance function.
什么是wavelength-by-wavelength product?
二者直接相乘
Figure2
Fig2（a）:
对于简单的乘积例子，给定高斯噪声时，后验分布的图像，最优解在岭处
Fig2（b）：
横截面表明，即使在岭处都有最大的固定值，一些局部区域也会含有不同的概率质量
什么是parametre vector？参数是什么意思？
Y=ax+b Y是因变量，x是自变量，a,b就是参数
什么是rendering equation？
https://zhuanlan.zhihu.com/p/52497510
什么是Gaussian observation noise？高斯噪声？
高斯噪声百度百科：https://baike.baidu.com/item/%E9%AB%98%E6%96%AF%E5%99%AA%E5%A3%B0/8587563?fr=aladdin
为什么深度学习去噪都采用高斯白噪声？https://www.zhihu.com/question/67938028
高斯白噪声解释：https://blog.csdn.net/szlcw1/article/details/41758711
Figure3
delta loss function损失函数？
常见损失函数：https://blog.csdn.net/perfect1t/article/details/88199179
应该就是指单峰函数，0-1函数$\delta(\widetilde{x}-x)=0\quad if(\widetilde{x}=x)$ 相同的是预判正确的，所以是没有损失的
期望损失？
https://blog.csdn.net/hx14301009/article/details/79870851
Introduction
A. Why Color Constancy is Difficult
1.Problem Statement
The entries of sj specify the fraction of incident light reflected in Nl evenly spaced wavelength bands throughout the visible spectrum
是不是错了，应该是reflected in $N_j$?这句话怎么理解?
应该意思是$N_l$个均匀排列波长带
2.Why It Is Difficult
It is underdetermined and it is nonlinear.
什么是欠定的？
http://blog.sina.com.cn/s/blog_531bb7630100xx6c.html
If we have data from N image locations (say, 10) and assume one illuminant, then we have NNr measurements (e.g., 10 x 3 = 30) available to estimate Nl(N + 1) scene parameters [e.g., 31 x (10 + 1) = 341].
这句话怎么理解？
To address the underdeterminancy of color constancy, previous investigators have described spectral functions by using low-dimensional linear models
什么是低维线性模型？
The columns of $B_e$ are the basis functions of the linear model, since the matrix product $B_ew_e$ expresses a weighted sum of these columns.
什么是基函数？
https://www.jianshu.com/p/5cc427f0df33
If we assume that a population of spectra lie within an $N_m$-dimensional linear model, then we can parameterize the spectra by specifying the model weights.
这句话如何理解？
