def _find_DoG_extrema(self, DoG):
# TODO: sample ?
ext = []
for octave in DoG:
_, rows, cols = octave.shape
# assert octave.shape = (<layers>(=s+2), <rows>, <columns>)
##########################
# time ?
##########################
peeled = [octave[ind1, ind2, ind3] for ind1, ind2, ind3 in
product(*[[slice(1, -1), slice(2, None), slice(None, -2)]]*3)]
center_block = peeled[0] # octave[1:-1,1:-1,1:-1], the center part
neighbor_blocks = peeled[1:] # neighbors in 26 directions in 3-D DoG space with offset 1
is_extreme = \
np.bitwise_or(center_block > (np.max(neighbor_blocks, axis=0)),
center_block < (np.min(neighbor_blocks, axis=0)))
# assert is_extreme.shape = (s, rows-2, columns-2)
ext_coord = np.array(
list(product(*[range(1, i-1) for i in octave.shape]))).reshape([x-2 for x in octave.shape]+[-1])
assert ext_coord.shape[:-1] == is_extreme.shape
ext_coord = ext_coord[is_extreme].astype(np.float)
print("%d key point candidates found" % ext_coord.shape[0])
# assert ext_coord.shape = (<number of key points>, 3)
ext_coord /= [1, rows, cols] # convert row, col coord to relative
ext_coord[:, 0] = get_sigma_by_layer(ext_coord[:, 0])
ext.extend(list(ext_coord))
# filters
# for layer, row, col in ext_coord:
# nb = octave[layer-1:layer+2, row-1:row+2, col-1:col+2] # 3x3x3 neighbor
# x_hat, d_x_hat = self._fit_extremum(nb)
# if
# TODO: if x[i] > 0.5, do it recursivey
# print(x_hat, d_x_hat)
return ext
评论列表
文章目录
