立体匹配文献阅读笔记（一）《A non-local cost aggregation method for stereo matching》12年

一、The most efficient local cost aggregation method is unnormalized box filtering which runs in linear time (relative to the number of image pixels) using integral image but blurs across depth edges. Edge-aware filters like bilateral filter are very effective for preserving depth edges. However,full-kernel implementation of the bilateral filter is slow.
这儿的 bilateral filter 不是指先对原始图像进行滤波处理，而是指在代价聚合的过程中“滤波”，可以理解为动态调整代价聚合。The two constants are usedto adjust the spatialsimilarity and the range (intensity/color) similarity respectively.

二、Recently,a new edge-aware filter called guided image filter is proposed. Unlike bilateral filter, its runtime is linear in the number of image pixels, and was demonstrated to outperform all the other local methods on Middlebury benchmark both in terms of speed and accuracy.

三、使用middlebury数据集实际测试表明，本文提出的NL代价聚合方法比现有的最好的局部代价聚合guided image filter的效果还要好。

四、本文提出的NL代价聚合可以用来对视差图进行后处理，即对检测出的误匹配区域即空洞区域进行填充，效果比一般局部后处理方法好。

五、关于怎么把全图像素点用MST最小生成树表示出来，需要找时间看下。