在上一篇中配置好Caffe for windows并配置matlab接口之后，我们尝试运行RCNN（区域卷积神经网络）的demo，在给出该方法的思想之前，我们尝试运行demo看看效果如何。此处特别感谢@kai提供指导。

一、下载代码与模型

首先在git for Rcnn下载RCNN的源代码，并在git for Selective Search下载(这个东西是什么我们接下来再说，先不要管他。)在ubuntu系统下运行 ./data/fetch_models.sh下载模型，总共大约1.5G。
全部下载后，将Selective Search复制到rcnn-master/selective-search，之后下载好的模型解压，得到的caffe_net和rcnn_models两个文件夹全部复制到windows系统下rcnn-master/data/。完毕

二、运行demo

打开matlab，将%caffe-root加入matlab的path中。
cd %rcnn-master 在命令行中输入 rcnn_build()来生成liblinear。结果如下：

Compiling the anisotropic gauss filtering of:
   J. Geusebroek, A. Smeulders, and J. van de Weijer
   Fast anisotropic gauss filtering
   IEEE Transactions on Image Processing, 2003
Source code/Project page:
   http://staff.science.uva.nl/~mark/downloads.html#anigauss
Compiling the segmentation algorithm of:
   P. Felzenszwalb and D. Huttenlocher
   Efficient Graph-Based Image Segmentation
   International Journal of Computer Vision, 2004
Source code/Project page:
   http://www.cs.brown.edu/~pff/segment/
Note: A small Matlab wrapper was made.
Compiling liblinear version 1.94
Source code page:
   http://www.csie.ntu.edu.tw/~cjlin/liblinear/

之后在命令行输入key=caffe(‘get_init_key’);以验证我们caffe和matlab的wrapper是否正确。
输入rcnn_demo查看，运行我们的RCNN_DEMO，显示如下

Welcome to the PASCAL demo
Running in GPU mode
(To run in CPU mode, call rcnn_demo(demo_choice, 0) instead)
Press any key to continue
Initializing R-CNN model (this might take a little while)
done
Computing candidate regions...found 2100 candidates (in 10.069s).
Extracting CNN features from regions...done (in 159.304s).
Scoring regions with detectors...done (in 1.704s)
Applying NMS...done (in 13.593s)
Total 20-class detection time: 192.384s

看上去很酷！