Adaptive point-cloud surface interpretation

We present a novel adaptive radial basis function network to reconstruct smooth closed surfaces and complete meshes from nonuniformly sampled noisy range data. The network is established using a heuristic learning strategy. Neurons can be inserted, removed or updated iteratively, adapting to the complexity and distribution of the underlying data. This flexibility is particularly suited to highly variable spatial frequencies, and is conducive to data compression with network representations. In addition, a greedy neighbourhood Extended Kalman Filter learning method is investigated, leading to a significant reduction of computational cost in the training process with desired prediction accuracy. Experimental results demonstrate the performance advantages of compact network representation for surface reconstruction from large amount of non-uniformly sampled incomplete point-clouds.