A key pose model for human interaction recognition and color from gray by optimized color ordering

In this thesis, we address two different problems in computer vision. First, the human interaction recognition is discussed whose goal is to recognize the action type of interacting humans in videos. We model the interaction using a sequence of key poses, important atomic-level actions. Spatial arrangements between the actors are included in the model as is strict temporal ordering of the key poses. Second, we attack the problem of color from gray. Our goal is to build a gray level image which can be used to recover color image with the minimum amount of embedded information. We use a parametric curve which maps the gray values to rich samples of color space. Next, we attach the parameters for the curve to the gray image in order to recover an approximation of color. The method rests on an optimization of parameters in which both the gray and color error are minimized.