An Arbitrary Point Tracking using Multi-scale Refined Optical Flow

This paper describes an algorithm for an arbitrary point tracking using multi-scale refined optical flow. Spatio-temporal approach used for optical flow estimation has been often criticized because of their limitations in handling large motions. We propose a novel method how to adapt it in a multi-scale refined strategy in this paper. In order to calculate optical flow which is sensitive to small and big motion, multi-scale representation is used and when combining two methods, Lucas-Kanade's optical flow estimation and multi-scale representation, refined method is used to prevent estimation error from increasing as the pyramid level goes down from top level. In addition, an accuracy of estimation is enhanced using optical flow constraints and also it makes possible to track points which has no characteristics of edge or corner using surrounded optical flow information of the point in an image. The proposed algorithm is implemented and primary results are shown in this paper.