Sphere-Meshes for Real-Time Hand Modeling and Tracking

摘要

Modern systems for real-time hand tracking rely on a combinationrnof discriminative and generative approaches to robustly recover handrnposes. Generative approaches require the specification of a geometricrnmodel. In this paper, we propose a the use of sphere-meshesrnas a novel geometric representation for real-time generative handrntracking. How tightly this model fits a specific user heavily affectsrntracking precision. We derive an optimization to non-rigidly deformrna template model to fit the user data in a number of poses. Thisrnoptimization jointly captures the user’s static and dynamic handrngeometry, thus facilitating high-precision registration. At the samerntime, the limited number of primitives in the tracking template allowsrnus to retain excellent computational performance. We confirmrnthis by embedding our models in an open source real-time registrationrnalgorithm to obtain a tracker steadily running at 60Hz. Werndemonstrate the effectiveness of our solution by qualitatively andrnquantitatively evaluating tracking precision on a variety of complexrnmotions. We show that the improved tracking accuracy at highrnframe-rate enables stable tracking of extended and complex motionrnsequences without the need for per-frame re-initialization. To enablernfurther research in the area of high-precision hand tracking, wernpublicly release source code and evaluation datasets.