Hough Forest---Hofsonin (Hough forest) target detection algorithm for target detection

Hough Forest target detection is a more fashionable target detection algorithm, Juergen Gall is proposed on 2009 CVPR.

Hough Forest sounds like a combination of Hough transformation +random forest, in fact, not exactly. It is more like the combination of decision forest and regression forest plus generalized Hough Transform: Each tree in the forest is not a classification tree or a regression tree, but each of these nodes may be a cluster node or a regression node. The classification node minimizes the Class-label uncertainty, while the regression node minimizes the offset uncertainty. Then output the 2D Hough Image (which can certainly be extended to 3d,4d), where the local extremum is the target parameter (position, scale, aspect ration).

Unlike the traditional target detectors, the training sample is P={i,c,d}, I is a local image block (patch) of the target, C is its class label, and D is its offset (offset) vector to the target Center (2D) The purpose of random tree growth is to minimize the class of node samples and offset impurity, respectively, corresponding to the decision node and the regression node. The split criterion on the node uses the method of pixel value comparison. The last leaf node records some statistics: CL represents a positive sample scale, and a set of DL sample offsets. In addition, the author also uses a cascade-like approach when training, forest=5trees+5trees+5trees, that is, after the first 5 trees have been grown, some of the more difficult samples are used to train the next batch of trees.

In the detection phase, extract the patch,pass down through every tree in the forest from the image, and then give the 2D Hough image a point on the x position to vote. The image is dense sampling and the final output is Hough image.

Here are some of the notes I read in my paper.

http://www.pamitc.org/cvpr13/program.php

http://blog.csdn.net/gxf1027/article/details/8308655

Hough Forest---Hofsonin (Hough forest) target detection algorithm for target detection