Adaboost tracking.
Group members:
Valentina Osipova; Ahmad Z. Mohammad.
To run the code:
`` cmake . make
./bin/nr1 face-model.xml img2.jpg
./bin/nr2 splice/splice.train splice/splice.test 50
./bin/nr3 nemo/frames.train nemo/frames.test 50 ``
Notes on the number of weak classifiers: as the number increases, the overall performance gets better, but it is very noisy as the stochastic search for a split value/attribute is different. However, the accuracy saturates at some point, and even shows some degradation after 20 Kindly check the attached graph.
For tracking nemo, Using non-overlapping negative examples enhanced the confidence drastically. Using overlapping negative examples confused the weak classifiers, and produced not-so-satisfactory results (unpredictable and inconsistent)