Online structured-output learning for real-time object tracking and detection

This demonstration will show two applications of online structured-output learning for real-time computer vision tasks. The first application is for 2D tracking of arbitrary objects, and uses an online kernelized structured-output SVM on a budget to provide robust adaptive tracking. The second application is for real-time 3D object detection. There has been much work recently in the computer vision community on highly efficient interest-point matching techniques, which allow real-time detection and pose estimation of objects. Existing methods store a database of interest-points extracted from reference views of the object, which are then matched in a nearest-neighbor fashion. We propose a different approach, which learns these descriptors discriminatively online in a structured-output SVM framework. This means the descriptors are optimized for detecting the target object in a given environment, providing improved detection performance. This work has been carried out in collaboration with Sony Computer Entertainment Europe (SCEE), and we have integrated our approach into a real-time simultaneous localization and mapping (SLAM) system running on the forthcoming Sony PlayStation Vita portable games console, where it provides the relocalization component for recovery from tracking failure.