3D Modeling from Multiple Images

Although the visual perception of 3D shape from 2D images is a basic capability of human beings, it remains challenging to computers. Hence, one goal of vision research is to computationally understand and model the latent 3D scene from the captured images, and provide human-like visual system for machines. In this paper, we present a method that is capable of building a realistic 3D model for the latent scene from multiple images taken at different viewpoints. Specifically, the reconstruction proceeds in two steps. First, generate dense depth map for each input image by a Bayesian-based inference model. Second, build a complete 3D model for the latent scene by integrating all reliable 3D information embedded in the depth maps. Experiments are conducted to demonstrate the effectiveness of the proposed approach.