Building a photo-realistic virtual world using view-dependent images and models

Describes a new approach for building a photorealistic virtual world based on the real world. Using an image capturing system mounted on a vehicle, we have assembled a vast image database with global position and orientation information. However, if we store all the images taken from various users' viewpoints, we might exhaust the resources. Thus, it is important to generate novel images from finite image data sets. For this purpose, we believe that a view-dependent "subjective" virtual space is sufficient, and an "objective" virtual space is not imperative. Based on this concept, we propose a view-dependent image projection method and a 3D layer presumption method to generate a subjective virtual space. By using these methods, we succeeded in constructing a photorealistic virtual world. Both methods yield only a local model of the entire virtual world. However, by changing the local models according to the user's viewpoint, the user can move around a wider area. In this paper, we demonstrate that this kind of subjective model is practical and sufficient to construct a wide photorealistic virtual world.