Pose estimation for textured and textureless objects
Dr. Manolis Lourakis
Foundation for Research and Technology - Hellas, Greece
Determining the 3D pose (i.e., position and orientation) of objects from images is a common requirement for vision systems applied to areas such as robotic manipulation, tracking, augmented reality, tangible interfaces, etc. Such systems should operate reliably in dynamic and unknown environments, delivering accurate object pose estimates despite any background clutter or variations in the appearance of objects due to changes in viewing position, illumination and occlusions. The most effective strategy for dealing with such challenges is to proceed according to the model-based paradigm, which involves building 3D models of objects and then determining object poses by fitting their models to images. The talk will describe techniques for estimating the pose of rigid objects from intensity and depth images. Ordinary intensity images are suitable for use when the objects of interest are textured. Object models in this case consist of sparse sets of 3D points and their associated local image features. When object surfaces lack texture, depth images provide information about object structure and 3D meshes or point clouds are used as object models.