Non-maximum suppression(非极大值抑制算法）

在RCNN系列目标检测中，有一个重要的算法，用于消除一些冗余的bounding box，这就是non-maximum suppression算法。

这里有一篇博客写的挺好的：

　　http://www.cnblogs.com/liekkas0626/p/5219244.html

借用博客里的两张图，如有问题，请联系我删除。

在目标检测中，这些bounding box都表示检测到了人脸，并且会给每一个bounding box一个score，最终我们需要只保留score最大的bounding box（记为bounding box1），將与bounding box1 overlap (重叠) 较大的一些bounding box消除，即只保留在这个局部区域score最高的一个bounding box。可能在一张图中有多个人脸，那么我们就需要保留各个局部区域score最大的bounding box，两个bounding box的重叠程度一般用IOU的值作为衡量，

IOU = 两个bounding box的围成的区域的交集 / 两个bounding box围成区域的并集。

 function pick = nms(boxes, overlap)
 % top = nms(boxes, overlap)
 % Non-maximum suppression. (FAST VERSION)
 % Greedily select high-scoring detections and skip detections
 % that are significantly covered by a previously selected
 % detection.
 %################################################################
 % input   -----    boxes: object detection window
 %                         size(boxes) = (n,)
 %                         n --- the number of window
 %                       (xMin, yMin, xMax, yMax, score)
 %                  overlap: suppression threshold
 %
 % output  -----    pick:the index of reserved window
 if isempty(boxes)
   pick = [];
   return;
 end

 x1 = boxes(:,);
 y1 = boxes(:,);
 x2 = boxes(:,);
 y2 = boxes(:,);
 s = boxes(:,end);

 % calculate the area of all detections
 area = (x2-x1+) .* (y2-y1+);
 % vals is the sorted elments in ascending order, I is the corresponding index
 [vals, I] = sort(s);

 pick = s*;
 counter = ;
 while ~isempty(I)
   last = length(I);
   i = I(last);
   pick(counter) = i;
   counter = counter + ;

   xx1 = max(x1(i), x1(I(:last-)));
   yy1 = max(y1(i), y1(I(:last-)));
   xx2 = min(x2(i), x2(I(:last-)));
   yy2 = min(y2(i), y2(I(:last-)));

   w = max(0.0, xx2-xx1+);
   h = max(0.0, yy2-yy1+);

   inter = w.*h;
   o = inter ./ (area(i) + area(I(:last-)) - inter);

   I = I(find(o<=overlap));
 end

 pick = pick(:(counter-));