<title>Pattern-vector-based reduction of 3D meshes created from multimodal data sets</title>

Abstract

Recent interest in photo-realistic 3D modeling has appeared in a variety of areas. These models are extremely useful in a realistic simulation environment. When viewing these large models, however, the frame-rate of even the fastest computers can be extremely low. Therefore, methods to lower the resolution of the model, and thus increase the frame-rate, while retaining the realism are desired. By using a multi-resolution model, low-resolution representations can be used while interacting with or moving the model, and the details from the higher resolution representation added when the model is stationary. To aid in the creation of such a model, we have developed a new triangle mesh reduction technique based on pattern vectors which creates a representation of a model. The reduction method presented here is based on the edge collapse/vertex split concept since this method lends itself easily to the multi-resolution mesh concept. The uniqueness of this reduction method, however, comes in the form of the data representation. We not only use the position of the vertices, but also other geometric properties such as normal and curvature. In addition, we have the advantage of having even more information about the mesh than just structural information. One of these comes in the form of grey-scale and/or color intensity images. Other information such as thermal or radiation data can also be added. The reduction is accomplished by mapping the data into a higher dimensional pattern space in which each vertex is represented by a pattern or feature vector. Each vertex is assigned a n-dimension pattern vector containing all the feature information. Using the initial network of edge connections and the newly created feature space, the order of the edge collapses can be determined using all the given information by calculating edge lengths in the higher dimensional space and getting an edge length order. Examples from a variety of photo-realistic multi-resolution models are given including models created from synthetic data, laser range images, and digital elevation maps.