2. Technical Issues
A. Polygonal Surface Reconstruction from Point Clouds
Fig A-1. Algorithm overview of polygonal surface reconstruction from point clouds.
This research proposes a new procedure for generating a polygonal surface for approximating arbitrary point cloud data with any topology.
Our intent is to directly reconstruct a subdivision surface from unorganized points.
The subdivision surface is well known to be a natural level-of-detail and memory-efficient representation of the polygonal surface.
Because we use the interpolatory subdivision method like a modified butterfly scheme, we don't need to sample the displacement of all vertices but just those of odd vertices for each level.
So displaced butterfly subdivision surfaces (DBSS) have more efficiency of memory than the ones by a Loop scheme.
left: point clouds, right: shrink-wrapped rough mesh
Fig A-2. Polygonal surface reconstruction from a Chinese scimitar model.
B. 3D Feature Detection
Fig B-1. Algorithm overview of 3D feature detection.
In the first step of this approach, we extract lines from object-space normal vectors.
We do not use shading information, since this represents a reduced level of geometric information.
However, normal vectors alone are not adequate when the surface changes abruptly which happens at the outlines of objects.
Therefore we also use a depth discontinuity filter.
The object-space normal vectors and depth information acquired in this first step are packed into textures.
In second step we apply edge detection to the textures, which yields lines.
Laplacian edge detection produces good results and is relatively easy to implement on modern graphics hardware.
We also use a simple Sobel filter to identify outlines reliably.
left: target model, right: our method
left: apparent ridges, right: suggestive contours
Fig B-2. Dragon model rendered with 3 different techniques.
Fig B-3. Toon shading application for various models.