Effective classifiers for detecting objects
Mayo M. (2007). Effective classifiers for detecting objects. In Proceedings of the Fourth International Conference on Computational Intelligence, Robotics, and Autonomous Systems (CIRAS ’07), Palmerston North, New Zealand.
Permanent Research Commons link: http://hdl.handle.net/10289/2171
Several state-of-the-art machine learning classifiers are compared for the purposes of object detection in complex images, using global image features derived from the Ohta color space and Local Binary Patterns. Image complexity in this sense refers to the degree to which the target objects are occluded and/or non-dominant (i.e. not in the foreground) in the image, and also the degree to which the images are cluttered with non-target objects. The results indicate that a voting ensemble of Support Vector Machines, Random Forests, and Boosted Decision Trees provide the best performance with AUC values of up to 0.92 and Equal Error Rate accuracies of up to 85.7% in stratified 10-fold cross validation experiments on the GRAZ02 complex image dataset.