A fast hybrid geomorphing LOD scheme

A significant amount of recent work on real-time multiresolution rendering of triangulated geometric models is built on the concept of iterative edge contraction(e.g. [Hoppe 1997]). The system we have developed is a hybrid method that integrates aspects of this kind of approach and of more classical discrete Level of Detail (LOD) techniques. Our goal was to produce a multiresolution rendering system that would produce minimal visual artifacts when rendering high resolution scene or object datasets produced from sensor data while maintaining the same graphic performance we get with optimized static versions of those models. The first step of the method is to partition the triangular mesh into a set of groups. The model is then globally decimated into a series of discrete LODs using an algorithm based on vertex pair contraction. Each discrete level is repartitioned along the same boundaries. Groups are shaped based on criteria such as compactness, orientation, texture, and desired granularity. At run-time, LOD levels and geomorph ratios between selected levels are computed for each group of the model, these groups are then rendered as precomputed triangle strips. Border points between groups are geomorphed in order to maintain seamless continuity between neighboring LOD groups at all time. To achieve this, we maintain a connexity graph between the various groups and apply a set of simple rules when morphing these points. Morphing is done for space and texture coordinates, normals and color. Adjusting the granularity of the multiresolution rendering through this grouping process provides us with multiple advantages. First, it allows us to extend smooth geomorphs over numerous frames by morphing more triangles over a longer distance rather than expanding/contracting the model by a few polygons every frame. Secondly, each group having a static topology, it can be efficiently rendered as a set of pre-computed triangle strips. These groups can also serve as the units of an anticipative paging algorithm for larger models. Finally, groups can be shaped so that they correspond to rectangular texture images, therefore optimizing texture memory usage and smoothing texture swapping for high resolution models.