3D Firework Reconstruction from a Given Videos

Reconstruction of a 3-dimension(3D) firework show from a given videos is a key technology in light source simulation in computer graphics, which can be more effective and real than traditional method. Although the firework model is already very mature, however, to our best knowledge, there is not any existing method that can reconstruct a firework show from a given video. And due to the lack of camera arguments and depth message, reconstruction is very challenging. In this paper, a method is proposed to solve the problem. A rendering model which requires some parameters which describe the color and position information of firework as input and generates a 3D firework show as output is constructed, and then the problem becomes getting the parameters needed for the rendering model from the given video. The parameters are divided into two groups according to the relevance, and then different neural networks including 3D Convolution Neural Network (3D-CNN) and Recurrent Neural Network(RNN) are designed respectively to extract these parameters needed by our rendering model from a given video. It is found to be practicable and effective to reconstruct a 3D firework from a given video by testing this work with some firework videos in various perspective.