Estimation of a large relative rotation between two images of a fast spinning marker-less golf ball

This paper proposes a method of rotation estimation of a fast spinning marker-less golf ball, which rotates up to 60 degrees in a 1-millisecond frame, from two consecutive frames of an infrared video sequence. The proposed method solely relies on the dimple pattern, which has slight irregularity in its configuration providing a clue for alignment, and does not need help of logo marks or fiducial markers on the ball surface. The method first narrows down relative pose candidates by registration of dimple center points, and then chooses the pose that enables best matching of dimple edge pixels. Experimental results show that it achieves 99 % success rate in 0.5 s computational time, which offers a practical solution for a golf simulator system.