Christian-A. Bohn

Surface Reconstruction with Smart Growing Cells

We propose Growing Cells Meshing (GCM) - a reconstruction algorithm which creates triangle meshes from clouds of arbitrary point samples registered on object surfaces. GCM is different to classical approaches in the way that it uses an artificial neural network together with an iterative learning technique to represent the triangle mesh. Based on the Growing Cell Structures (GCS) approach [3] we introduce the Smart Growing Cells (SGC) network as extension to fulfill the requirements of surface reconstruction. Our method profits from the well-know benefits entailed by neural networks, like autonomy, robustness, scalability, the ability of retrieving information from very complex data, and adaptability. On the downside, typical drawbacks like undesirable smoothing of information, inability to exactly model detailed, discontinuous data, or a vast amount of computing resources at big network sizes are overcome for the application of surface reconstruction. The GCM approach creates high-quality triangulations of billions of points in few minutes. It perfectly covers any amount and distribution of samples, holes, and inconsistent data. It discovers and represents edges, manages clusters of input sample points, and it is capable of dynamically adapting to incremental sample data.

Growing Surface Structures: A Topology Focused Learning Scheme

Iterative refinement approaches derived from unsupervised artificial neural network (ANN) methods, such as Growing Cell Structures (GCS), have proven very efficient for the application of surface reconstruction from scattered 3D points. The Growing Surface Structures (GSS) algorithm is a major conceptual change in the GCS approach. Instead of “adjusting” the learning behavior, the central learning scheme is shifted from optimizing the distribution of vertices to the creation of a valid surface model. Where in former GCS approaches the created topology is only implicitly represented in the process, it is explicitly integrated and represented in the refinement process of the GSS approach. Here the closest surface structure, such as a vertex, an edge or a triangle is found for a given sample and the actual sample -to-surface distance is measured. With this additional information the adaptation process can be focused on the created topology. We demonstrate the performance of the novel concept in the area of surface reconstruction.

Smart Growing Cells: Supervising Unsupervised Learning

In many cases it is reasonable to augment general unsupervised learning by additional algorithmic structures. Kohonens self-organzing map is a typical example for such kinds of approaches. Here a 2D mesh is superimposed on pure unsupervised learning to extract topological relationships from the training data. In this work, we propose generalizing the idea of application-focused modification of ideal, unsupervised learning by the development of the smart growing cells (SGC) based on Fritzke’s growing cells structures (GCS). We substantiate this idea by presenting an algorithm which solves the well-known problem of surface reconstruction based on 3D point clouds and which outperforms the most classical approaches concerning quality and robustness.