codesfarmer / stnumber Goto Github PK

This project include some interface to help the users using caffe more convenient

• Caffe
• HDF5
• tinyXML2
• OpenCV

After the depdendencies are setted up, you can compile and install our tools
#You are supposed under the root directory of our project
$mkdir build && cd build
$cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/where/you/want/install ..
$make install

1.The program can generate bounding boxes with PNet, RNet, ONet of MTCNN

To use the program, you should prepare your images and annotations in the follow struct

---------------folder1-----------whateveryouwant--------------------cam0
　　　　　　|　　　　　　　　　　　　　　　　　|------xml
　　　　　　|
　　　　　folder2-----------whateveryouwant--------------------cam0
　　　　　　|　　　　　　　　　　　　　　　　　|------xml
　　　　　　|
　　　　　　.
　　　　　　.
　　　　　　.
Then a filelists shuold be generated, which include the full path of the folder include the cam0 and xml, and the file name followed.
For example, filelists.txt, like this
/root/pathtofolder1/folder1/whateveryouwant/beforecam0 image_0001
/root/pathtofolder1/folder1/whateveryouwant/beforecam0 image_0002
...

To generate patches, you can run the
patchtest filelists.txt path/to/save/generated/samples
You can refer examples in test/test_patch.cc for more details

2.The tools also support write the data into HDF5 database, you can refer the test/test_hdf5.cc to create your HDF5 file

3.You can extract the output from any layer of convolutional neural network(examples at test/test_cnn.cc)

For *net_ir1.caffemodel and *net_ir2.caffemodel model, we employ different mean value for normalization
The mean value for ir1 is 17.2196, while for ir2 it is 0, the down scales are same.