Large Histogram Computation for Normalized Mutual Information on GPU
Our two-step image re-alignment uses NMI metric. To overcome limitations e.g. computing single coefficient, approximating entropy or reducing the histogram bin number, we developed 2 CUDA kernels. One considers non-valid pixels the other a multitude of problems. Both compute all the NMI coefficients using 65536-bin histogram allowing exhaustive searches of the best registration or keypoint pairs.