View interpolation using neural network for fullsphere 3D telepresence

As the Internet and video equipment develop, service using VR technology will become popular. Telepresence use those technologies. This is expected to be applied in various fields, e.g. tourism. Multiple users can use the telepresence system at the same time by using the full sphere(Omnidirectional) camera. However, stereoscopic vision is impaired, because the camera is fixed at the time of recording and cannot respond to the position movement of the user's point of view. This paper aims to preserve stereoscopic vision in the fullsphere telepresence system by estimating the image of the position of the viewpoint of the user from the existing fullsphere image by the neural network. Images of the current line of sight direction is taken out from fullsphere cameras arranged in concentric circles and the image of the midpoint is estimated from the images of the two viewpoints by using the neural network. By using the image obtained at the midpoint, this processing is performed recursively to retrieve the image of the viewpoint of the user. In this paper, we experimentally verified whether the midpoint viewpoint can be estimated by neural network.