nms_maxpool(self, R, k, ksize): """ Get top k interest points that are local maxima over (ksize,ksize) neighborhood. One simple way to do non-maximum suppression is to simply pick a local maximum over some window size (u, v). Note that this would give us all local maxima even when they have a really low score compare to other local maxima. It might be useful to threshold out low value score before doing the pooling. Threshold globally everything below the median to zero, and then MaxPool over a 7x7 kernel. This will fill every entry in the subgrids with the maximum nearby value. Binarize the image according to locations that are equal to their maximum. Multiply this binary image, multiplied with the cornerness response values