Top-down 3D Tracking and Pose Estimation of a Die Using Check-points

Abstract

Real-time 3D pose estimation from monocular image sequences is a challenging research topic. Although current methods are able to recover 3D pose, they require a high computational cost to process high-resolution images in a video sequence at high frame-rates. To address that problem, we introduce the new concept of check-points. They are the minimum number of points needed to detect a 3D object motion. Our method tracks the 2D projections of the check-points over a 2D maximum pyramid. To handle large displacements of the object, our approach evaluates the projection of the check-points at highest levels of the pyramid. Moreover, it refines the pose localization by utilizing the check-points at lowest levels of the hierarchy. We show that just checking a few cells per frame, our method estimates the 3D pose of the tracked object. This early version of the method works with a specific type of object a 3D cube, with six well distinguished faces and which salient features in all the faces are dots, a die.