Abstract
With the development of human pose estimation technology from an image/video, 2D skeleton pose becomes ubiquitously available. Different from the traditional 3D poses, it is more challenging to understand 2D poses such as action recognition because joint positions in 2D poses change a lot when the camera changes its capturing angle. Thus, it is essential to generate a front view pose from a non-front view one to avoid the camera angle problem. In this paper, we propose a novel algorithm for front view generation of 2D poses using homography transformation in an adaptive window. Our experiments show that the algorithm can greatly reduce the errors when computing the parameters of homography transformation.
