Robot Vision Assignment 2b/3a

1. Create a directory called c:\AdaBoost.
2. Download the following file containing sample database: http://www.cs.okstate.edu/~katchou/AdaBoost.zip
3. AdaBoost.zip contains positive.zip and scenery.zip. Unpack contents of positive.zip to c:\AdaBoost\positive\. Unpack scenery.zip over to c:\AdaBoost\scenery\
4. The rest of the instructions assume that you are using Microsoft Visual Studio versions 6 and higher. Create the following 2 Win32 Console projects: c:\AdaBoost\vboost and c:\AdaBoost\vdetect. Both should initially be empty.
5. Create the following subdirectories: c:\AdaBoost\vboost\classifiers c:\AdaBoost\vboost\weights and c:\AdaBoost\vboost\histograms
6. Download vboost.cpp to c:\AdaBoost\vboost and add to vboost project.
7. Download vdetect-scale.cpp to c:\AdaBoost\vdetect and add to vdetect project. Make sure both projects compile.
8. Run vboost to create a team of 100 classifiers. It is recommended to use at least 500 positive and 500 negative samples, the more the better: vboost new.cls 500 500 1 100. This takes about 20 minutes on a 3GHz P4 computer.
9. Optionally, run vboost new.cls 500 500 1 100 again to produce a classifier file in text format and view it in text editor. The file name will be new.txt
10. Copy new.cls and/or new.txt over to c:\AdaBoost\vdetect
11. Download some group photograps in pgm format to c:\adaboost\vdetect
12. Run detector program on each of them: vdetect.exe -i new.cls <photo name>. The following output will be produced:
    • 1) scale.pgm - a scaled version of your image.
    • 2)detected.pgm - a file showing detection results.
13. The following parameters affect performance of your detection:
    • 1) vdetect-scale.cpp line 893, classifier.thresh = 0.62 - try to experiment with threshold  values from 0.4 to 0.8 and observe results.  
    • 2) vdetect-scale.cpp line 411, double scale = 1.1; - try to experiment with scale values from 0.8 to 1.4 and observe results.
14. Once you are satisfied with results, go back to c:\adaboost\vboost and remove or rename new.cls. Your next task is to create a team of classifiers that only uses 3 different types of features.
15. Go back to vboost.cpp and modify int generate_features_18k(int max_h,int max_w)  to produce only 3 types of features. This can be achieved by commenting out one of the loop blocks starting with type =. Repeat steps 7-13 and observe results. Good luck.