ABSTRACT

Since the human hand is kinematically complex with 22-30 degrees of freedom with prevalent self-occlusion we resort to a strong geometric proxy to gain robustness. Our approach faithfully approximates both shape and shading of a hand in an unseen target pose even with large unobservable hand parts in the prototype images used for synthesizing. This allows the use of less generalizing, pixel-wise objective functions that incorporate shading information.