23. Large Histogram Computation for Normalized Mutual Information on GPU
Authors: Sophie Voisin (Oak Ridge National Laboratory)Devin A. White (Oak Ridge National Laboratory)Jeremy S. Archuleta (Oak Ridge National Laboratory)
Abstract: Our Photogrammetric Registration of Imagery from Manned and Unmanned Systems pipeline requires accurate computation of multiple Normalized Mutual Information (NMI) coefficients to perform a registration process in two steps, referred as global localization and registration refinement. Computing the NMI coefficients requires generating large joint-histograms to compute the images’ joint-entropy, which allows performing an exhaustive search of all possible translations between two images for global localization, and to perform accurate keypoint matching for registration refinement.
Contrary to existing GPU implementation our kernels uniquely accommodate the specificity of each step. They both compute all the NMI coefficients at once using 65536-bins joint-histograms. However the first kernel has the particularity to use a mask of valid pixels for the computation while the second kernel massively compute a multitude of small registration problem corresponding to each keypoint matching with intensity values as descriptor.
The poster will provide implementation details and performance for both kernels.
Two-page extended abstract: pdf