Data Mining from Multimedia and Human Behavioral Data

Object recognition/detection has always been a challenging topic for both computer scientists and psychologists. We study the categorical object recognition/detection problem from views of both computer vision and human vision, with the focus on feature representation. We propose representing general object categories using multiple types of complementary features, and using AdaBoost to select the most discriminative features. We also present a computational model which describes human eye movements during a visual search task. We then incorporate our heterogeneous-feature representation into the human eye movements model, and further explore human search behavior when the search target is categorical. The model is shown to have strong agreement with human behavior. Motivated by the search behavior of humans, we propose to use multi-resolution as a general framework for object category detection which achieves real-time detection speed with fast training and high accuracy. We also explore the concept of "visual similarity" in humans from the point of view of "nontargets".