dtosato/arco

ArCO is multi-class object detector based on a patch based model able to deal with low resolution objects.

ArCO

Disclaimer

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. Feel free to use this code for academic purposes. Please use the citation provided below.

The test part of this code takes 0.05 seconds per image on a Intel Xeon(R) CPU E5440 @2.83 GHz 8 GB RAM on Matlab 2009b. The most computationally expensive part of this code is the learning phase Y_train_light.m

Performance

This code performs patch-based head pose detection given the a 50x50 image. Should you have any problems please email me at [email protected]

Data

• QMUL4PoseHeads

Demo code and images provides

There is one complete script for learning and testing ARCO:

1. Download data.
2. Z_ARCO.m: this is the main script. It is able to learn and test a multi-class Logitboost classifier on Riemannian Manifold.
3. The variable 'experiment' (in Z_ARCO.m) contains a path where all the pre-computed parts of this framework are stored. Only the classification results are not computed in order to show you some qualitative results of this framework.
4. If you want to test this framework on the complete test set, just change the variable 'test_dir' from './QML4PoseHeads/test_demo' to './QML4PoseHeads/test'.
5. If you want to see the statistics of this framework on the complete test set without run testing, these are in [experiment '/full_test_results'].

This code is provided with a pre-computed training set and its learned classifier in order to directly test the classifier.

Citation

Diego Tosato, Michela Farenzena, Mauro Spera, Marco Cristani, Vittorio Murino “Multi-class Classification on Riemannian Manifolds for Video Surveillance,” ECCV, 2010.