Introduction

Explosive growth in computing power has realized the potential for 3D graphics programs and devices. Increased demand for 3D computer graphics is being fueled by industries such as Hollywood movies, television broadcasting and commercials, special effects based ride films, the Internet, and the PC and console gaming industries. Computer graphics software applications that facilitate the development of products in these industries all require the use of 3D models. We therefore set upon ourselves the task of finding a quick and easy way to generate 3D models that requires minimal human intervention.

Our program will synthesize a 3D texture-mapped model from a sequence of 2D images of an object taken with a digital camera or extracted from an MPEG stream. Once this set of 2D images has been loaded into a computer, our collection of algorithms, the Geometry Inference Engine (GIE), will output a three-dimensional data set. Essentially, our GIE engine uses a technique called Image Factorization, which extracts geometric information from the pictures and reconstructs the shape of the real-life object in the virtual environment. Our engine also extracts the color information from the digital photographs and uses the information to texture map the polygon mesh. Finally, we will use 3D view synthesis to guess the shape of any hidden parts of the object. The result is one seamless virtual environment with a beautifully modeled 3D object in the center. This technique is faster and cheaper than traditional techniques such as laser scanning and by-hand modeling.

Why there is a Need

Gaming companies build models in-house at an extraordinary cost. According to Richard Albritton of CG², it costs his company $1600 and 40 hours to develop a single model!. At Ritual Software, Charles Brown usually sees "12 [models] out of the fast modelers and 7 out of the slower ones" each month. This weak output level demonstrates the opportunity and need for our technology. This assessment is also evidenced by Albritton’s statement that "some customers have given us lists of 30-50 models that they need yesterday, so if we could produce them faster, that would be an asset."

In efforts to boost output, companies have tried utilizing laser scanners. However, this technology can not easily be fitted to this market, and is therefore inadequate. For example, Albritton attempted to use laser scanning for a real military tank; the project was a failure. Setup and staging for scanning the vehicle took three months, as it had to be cleaned with a special solution. Scanning took several days, and resulted in a data set being much too detailed for today’s (and tomorrow’s) hardware. "There were practically no computers that could do much with the data other than just load it." Reducing the polygon count was costly and time consuming.

Our project fits in nicely with their primary requirements - low polygon count meshes of real world objects with textures mapped accurately onto it. And so we begin our journey into the innards of the Geometry Inference Engine.

[ Back ] [ Home ] [ Next ]