Abstract:
Animating 3D character models has attracted the interest of many researchers over
the past two decades, and many related practical algorithms have been developed.
These algorithms apply various techniques ranging from physics-based animation
and inverse kinematics to 3D skeletal animation and rigging. In this study, we
demonstrate a framework to reconstruct the 3D models of the players in sport
game views. Initially, pose features are extracted from each player’s body using a
set of deep neural networks. These networks are pre-trained on 3D player data.
Next, these poses are applied to the 3D model of the player. Eventually, The poses
and positions of the players in the virtual field will match the actual ones. The
view can be displayed via a 3D viewer or virtual reality devices. Our proposed
framework consists of different deep neural networks, including convolutional and,
recurrent neural networks, for estimating human poses that form the main body.
In this study, we were able to reconstruct the body models of real players and
transform them into avatars. In addition, we outperformed the rendering process
of an existing research.
