文摘
Proposes hand pose estimation using a combination of model optimisation and discriminative methods which allows tracking to be performed at over 40 frames per second using a single CPU thread. Introduces a residual error regression for hand pose estimation, learning from mistakes in model optimisation. A method of training, which captures system response and user variance, allowing supervised feedback for joint refinement. Extensive quantitative and qualitative evaluation including additional datasets and comparison against multiple state of the art approaches.