Segmentation obtained by thresholding the UCM at certain level

这篇主要讲如何获得最右边的结果，论文可以参考：
From Contours to Regions: An Empirical Evaluation
Contour Detection and Hierachical Image Segmentation

根据上图，我们可以得到一般的思路，给定原图和其对应的UCM以及某个参数因子，就可以得到我们想要的结果。大致步骤如下：
1、读取某一原图及其对应的UCM
2、设定thresholding
3、根据thresholding，得到某一level的UCM map
4、对该UCM map的所有regions设定label
5、构建一个三通道的RGB矩阵，每个通道都是thresholding对应level的UCM map
6、找到所有UCM map 的label在原图对应的二维坐标
7、在原图基础上，求regions的label对应坐标位置三个通道的像素值，然后求average
8、对RGB对应的三个通道的每个regions赋值相应的average

代码如下：

load('data/101087_ucm2.mat','ucm2');image=imread('data/101087.jpg');%convert ucm to the size of the original imageucm = ucm2(3:2:end, 3:2:end);%get the boundaries of segmentation at scale k in range [0 1]k = 0.4;bdry = (ucm >= k);%get superpixels at scale k without boundaries:labels2 = bwlabel(ucm2 <= k);labels = labels2(2:2:end, 2:2:end);% figure;imshow('data/101087.jpg');% figure;imshow(ucm);% figure;imshow(bdry);% change formattemp=mat2gray(bdry);temp=uint8(temp*255);%set a three channels matrix to save result,all the channel are%identical,etc,temp. The three channel,as we all know, reprents RGBRGB(:,:,1)=temp;RGB(:,:,2)=temp;RGB(:,:,3)=temp;% the number of different label values in labels matrixlabel=unique(labels);for i=1:length(label)    [x,y]=find(labels==i); % find the location where labels equal i    [m,n]=size([x,y]); %m is the numberi of piexl node of the region    sum=zeros(3,1); % make a 3x1 matrix to save the sum value    for j=1:m  % calculate the sum        sum(1,1)=sum(1,1)+double(image(x(j),y(j),1));        sum(2,1)=sum(2,1)+double(image(x(j),y(j),2));        sum(3,1)=sum(3,1)+double(image(x(j),y(j),3));    end    %calculate the average    average1=round(sum(1,1)/m);    average2=round(sum(2,1)/m);    average3=round(sum(3,1)/m);    for k=1:m        % set each channel         RGB(x(k),y(k),1)=uint8(average1);        RGB(x(k),y(k),2)=uint8(average2);        RGB(x(k),y(k),3)=uint8(average3);    endend%if you want show the contour,you should need the following code[x2,y2]=find(temp==255);[m2,n2]=size([x2,y2]);for i=1:m2    %set contour for each channel    RGB(x2(i),y2(i),1)=255;    RGB(x2(i),y2(i),2)=255;    RGB(x2(i),y2(i),3)=255;end

该代码中读取原图和UCM，是直接利用伯克利提供的源码BSR的example，可以做相应的修改。
直观图片数据如下：

原图

UCM

某一level的UCM map

没有边界轮廓的结果

有边界轮廓的结果