The 3D model of the person is obtained using depth from stereo. The range information obtained from depth from stereo is used to segment the background. Using a 3D meshing algorithm the surface of the 3D avatar is reconstructed from the point cloud data of the foreground. Since the model of the avatar is constructed in real-time the 3D shape or size of the person need not be known. Thus this method can be used to generate more realistic video avatars.