Transparent-supported radiance regression function

A modified RRF[Ren et al. 2013] rendering method called TsRRF is presented in this paper, which support global illumination in realtime for scenes with moving transparent objects. The key idea of this method is to augment the map between object and scene. There are two kinds of method to augment this map. First, we choose different attributes which can represent the true color of an object and the relationship with the whole scene in space, and at the same time in order to get these attributes, we use GPGPU to get real time information. Second we use deep learning to get the most important information from the sample data which can decrease the overfitting. In order to get more details and make full use of the sample data, we not only partition the scene by position, but also partition by object, and we will use different TsRRF to render different light effect like reflection or refraction. The network forward propagate process will also be put into the GPU and use the parallel feature to calculate quickly. As a result, the modified method works well when dealing with the transparent objects and have a real time effect.