Active Basis C++ Code

QuickStart| Implementation | Usage | Consistency | Screen Shot

Version problem: The package compiles in CMake 2.4 and VC 9.0. If you use CMake 2.6, some workaround is needed. If you use an older version of VC, a full version of boost is needed.

1. QuickStart

Follow the following steps to start playing

Download the code at active_basis_release_1, unzip it to some folder (say active_basis_cpp_dir).
Download and install CMake.
Open CMake,
- specify 'active_basis_cpp_dir\build\' as binary directory,
- spceify 'active_basis_cpp_dir\' as source directory,
- then click 'Configure', then wait a minute and click 'Configure' again, then click 'OK'.
- This will generate VC project files in 'active_basis_cpp_dir\build' folder.
Open 'active_basis_cpp_dir\build\ACTIVE_BASIS_Prj.sln', switch to release configuration and compile all projects.
[Exp1] Go to folder 'active_basis_cpp_dir\data\ex1_car' and
- run tilt.bat --- This will generate 'nlf_*' files in 'model' folder, which contain the natural image ensemble statistics.
- run ab.bat --- This will generate 'learned_sketches_*' files in 'model' folder, which contain learned sketches.
- run vis_active_basis.bat --- This will read the visualization file generated by ab.bat in folder 'tmp' and visulize them. Press enter on the visulization window to continue.
- open param.ini and modify parameters to try again.
[Exp3] Go to folder 'active_basis_cpp_dir\data\ex3_horse' and
- run tilt.bat
- run ab.bat
- run vis_active_basis.bat to visulize, optional
- run abtest.bat --- This will test the learned model and draw ROC.
- open param.ini and modify parameters to try again.
[Exp4] Go to folder 'active_basis_cpp_dir\data\ex4_hs_hrs_butf' and
- run tilt.bat
- run abem.bat --- This will use EM algorithm to do clusting. This will generate 'learned_EM_* files' in '..\..\data\model' folder, which contain learned EM model.
- run vis_active_basis_em.bat to visulize, optional
- open param.ini and modify parameters to try again.

2. Implementation

I use the following libraries, and pack them together in the compressed package, so you need not to download them yourself.

OpenCV --- reading image, filtering.
Boost::ublas --- matrices and vectors.
Boost::program_options --- command line parsing and configuration file parsing.
PlPlot --- visualization. If you want to debug the project containing PlPlot, you need download and install it.

I use CMake to generate the project files. CMake is a very small and neat tool.
I convert all double data to float, which results in nearly no accuracy loss but reduces computation and storage.

Open the project, you will find 6 main projects:

tilt.exe --- Corresponde to tilt.m+Chistog.c in Matlab code.
active_basis.lib --- A library project packs all main algorithms for all the followed projects.
active_basis_learn.exe --- Corresponde to ABlearn.m+Clearn.c in Matlab code.
active_basis_test.exe --- Corresponde to ABtest.m+ABtst_one.m+Ctesting.c in Matlab code.
active_basis_em_learn.exe --- Corresponde to ABlearn.m+Clearn.c in EM version of Matlab code.
active_basis_em_test.exe --- Testing the ROC of EM model, No corresponding Matlab code.

The project also contains 3 visulization projects:

vis_active_basis.exe
vis_active_basis_em.exe
vis_active_basis_em_test.exe

Additionally, it also contains 3 support projects:

dir_utils.lib --- Warper of Boost::program_options for command-line options.
stat_utils.lib --- Histogram calculation, ROC computing etc.
ublas_utils.lib --- Several input and output functions related to boost::ublas.
prog_opt.lib --- Boost::program_options, I pack it into the project for self-containing.

3. Usage

How to debug in VC:

Switch to Debug configuration and compile all.
Select the interested project and set as StartUp project.
View project property, go 'Debuging'.
Set the command arguments, referring to the '.bat' file in the 'data' folder.
Set the working directory as 'active_basis_cpp_dir\bin\debug'.

The command-line options. All executables accepts several of the following options

--train_pos_dir --- folder of training positive examples
--train_neg_dir --- folder of training negative examples
--test_pos_dir --- folder of testing positive examples
--test_neg_dir --- folder of testing negative examples
--temp_data_dir --- folder of temporary data (e.g. visualization)
--param_file --- full path and name of the parameter file
--model_dir --- folder of learned model

An example:
active_basis_em_test --test_pos_dir=%ddir%\posTest --test_neg_dir=%ddir%\negTest --param_file=%ddir%\param.ini --model_dir=%ddir%\model\

4. Consistency with Matlab Code

I test the code on 'Head-Shoulder' data and 'Header-Shoulder-Horse-Butterfly' mixtrain data.

For single template version, if I import the filtering results from matlab, the learned templates and final ROC will be exactly the same.
For EM version, if I import the filtering results from matlab, the learned templates may be have minor differences due to double -float conversion, but C++ version has higher coding gain. For each template, there may exist one sketch has 2 pixels deviation.
For single template version, if I use OpenCV filtering, the learned templates scores and final ROC will have minor differences.
For EM testing, there is no Matlab code for comparison.

5. Screen Shot

PlPlot is really a good tool for scientific plotting, I recomend everyone to use it.

Tilt

Learn

Test

Learn EM

QuickStart| Implementation | Usage | Consistency | Screen Shot

Haifeng GONG