Gpu­based Multi­view Rendering 2. Previous Works

Stereoscopic images in computer graphics applications often require two rendering passes reducing by half the frame rate. In this situation, the conversion from standard to stereoscopic images may involve some difficulties to maintain real time rendering if the geometry is made of thousands triangles. Since few years, auto-stereoscopic displays have become more and more popular because of their multiuser capability and because they do not require any specific glasses. However, they usually require five or more input views that can be difficult to generate in real time. In this paper, we present a single pass algorithm using GPU that speeds-up the rendering of stereoscopic and multi-view images. The geometry is duplicated using a shader program that reduces the data transfer between the main memory and the graphic card. It also brings together the computation of some vertices's properties that are similar from one view to another. 1. INTRODUCTION Stereoscopy is a technique that enables to watch three-dimensional images on 2D display thanks, most of the time, to specific glasses. It has many applications in various fields such as data visualization, virtual reality or entertainment because it tends to reproduce our visual perception and makes information easier to understand. In computer graphics, stereoscopic rendering consists in generating two images of a virtual environment from two slightly different viewpoints. In other words, it requires to render the geometry of the scene twice which can double the computational time. In such case, it can be difficult to maintain real-time rendering especially for applications like video-games that are complex in term of geometry and visual effects. Auto-stereoscopy is a technology recently applied to LCD displays [2] that introduces the ability for one or several users to watch stereoscopic images without wearing any glasses. Depending on their characteristics, auto-stereoscopic displays require from 5 to 64 images [9] to display a single 3D frame. A filter, made of small lenses, is overlaid on the surface of the screen and ensures to emit each image in a specific direction. So, if the user is well located in front of the display, each eye can see a single specific image. However, the important number of required input images makes it more difficult to maintain real-time rendering compared to a single view rendering. We can state two facts in term of time for standard multi-view rendering. Firstly, data transfer from main memory to the graphic card is costly. Secondly, …