Incremental discovery of object parts in video sequences

摘要

This paper addresses the problem of automatically discovering the rigid parts of an initially unknown moving deformable object in a monocular video sequence. The parts are first extracted through motion-based segmentation, using a time scale automatically chosen with the quantity of motion concept. Tracking and reobservation reinforce these low-level segmentation results and further segmentation is performed only when and where no modeled parts can be tracked. Central to the system is the Modeler that minimizes the impacts of erroneous segmentations and departures in tracking. The sequential nature of the framework allows incremental modeling and segmentation of parts that need not simultaneously be visible or in motion, making it possible to circumvent the typical constraint of model initialization. The fundamental principles are strictly ensemblist and do not rely on any specific PDF. The interest of this framework is demonstrated on three types of video sequences including human and robot motion.