def test_grid():
""" """
annot = np.zeros((100, 100))
annot[:, 60:] = 2
annot[15:65, 35:85] = 1
np.random.seed(0)
noise = annot + np.random.randn(*annot.shape)
unary = np.tile(noise[:, :, np.newaxis], [1, 1, 3])
tmp = (unary[:, :, 1] - 1)
tmp[annot == 0] *= -1
unary[:, :, 1] = tmp
unary[:, :, 2] = 2 - unary[:, :, 2]
fig, axarr = plt.subplots(ncols=unary.shape[-1],
figsize=(unary.shape[-1] * PLOT_SIZE, PLOT_SIZE))
draw_unary(axarr, unary)
fig.tight_layout()
fig.savefig(os.path.join(DIR_IMAGES, 'grid_unary.png'))
pairwise = (1 - np.eye(3)) * 10
labels = gco.cut_grid_graph_simple(unary, pairwise, n_iter=-1)
fig, axarr = plt.subplots(ncols=2, figsize=(2 * PLOT_SIZE, PLOT_SIZE))
axarr[0].set_title('original annotation')
axarr[0].imshow(annot, interpolation="nearest")
axarr[1].set_title('resulting labeling')
axarr[1].imshow(labels.reshape(*annot.shape), interpolation="nearest")
axarr[1].contour(annot, colors='w')
fig.tight_layout()
fig.savefig(os.path.join(DIR_IMAGES, 'grid_labels.png')), plt.close()
def estimate_atlas_graphcut_simple(imgs, ptn_weights, coef=1.):
""" run the graphcut to estimate atlas from computed unary terms
source: https://github.com/yujiali/pyGCO
:param list(ndarray) imgs: list of input binary images [np.array<height, width>]
:param ndarray ptn_weights: binary ptn selection np.array<nb_imgs, nb_lbs>
:param float coef: coefficient for graphcut
:return list(int):
>>> atlas = np.zeros((8, 12), dtype=int)
>>> atlas[:3, 1:5] = 1
>>> atlas[3:7, 6:12] = 2
>>> luts = np.array([[0, 1, 0]] * 3 + [[0, 0, 1]] * 3 + [[0, 1, 1]] * 3)
>>> imgs = [lut[atlas] for lut in luts]
>>> estimate_atlas_graphcut_simple(imgs, luts[:, 1:]).astype(int)
array([[0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
[0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
[0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1],
[0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1],
[0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1],
[0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1],
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]])
>>> np.sum(abs(estimate_atlas_graphcut_simple(imgs, luts[:, :1]).astype(int)))
0
"""
logging.debug('estimate atlas via GraphCut from Potts model')
if ptn_weights.shape[1] <= 1:
logging.warning('nothing to do for single label')
labels = np.zeros(imgs[0].shape)
return labels
labeling_sum = compute_positive_cost_images_weights(imgs, ptn_weights)
unary_cost = np.array(-1 * labeling_sum, dtype=np.int32)
logging.debug('graph unaries potentials %r: \n %r', unary_cost.shape,
list(zip(np.histogram(unary_cost, bins=10))))
# original and the right way..
pairwise = (1 - np.eye(labeling_sum.shape[-1])) * coef
pairwise_cost = np.array(pairwise, dtype=np.int32)
logging.debug('graph pairwise coefs %r', pairwise_cost.shape)
# run GraphCut
try:
labels = cut_grid_graph_simple(unary_cost, pairwise_cost,
algorithm='expansion')
except Exception:
logging.exception('cut_grid_graph_simple')
labels = np.argmin(unary_cost, axis=1)
# reshape labels
labels = labels.reshape(labeling_sum.shape[:2])
logging.debug('resulting labelling %r: \n %r', labels.shape, labels)
return labels
def perform_binary_cut(foreground, background):
"""Perform graph-cut based segmentation.
Parameters
----------
foreground, background: float32
Foreground, background probability matrix (w*h*3)
"""
eps = np.finfo("float").eps
foreground_matrix = -np.log(foreground + eps) / np.log(eps)
background_matrix = -np.log(background + eps) / np.log(eps)
stacked = np.dstack((foreground_matrix, background_matrix))
pairwise = 1 - np.eye(2)
print(stacked.shape, pairwise.shape)
segmentation = gco.cut_grid_graph_simple(stacked,
pairwise,
n_iter=-1,
algorithm="expansion")
return segmentation.reshape(foreground.shape)
def test_binary():
""" """
annot = np.zeros((100, 100))
annot[20:70, 30:80] = 1
np.random.seed(0)
img = np.random.randn(*annot.shape)
img += 2 * annot - 1
# !!! Be careful when doing this concatenation,
# it seems 'c_' does not create a copy
# u = np.c_[img.flatten().copy(), - img.flatten().copy()]
unary = np.tile(img[:, :, np.newaxis], [1, 1, 2])
unary[:, :, 0] = img
unary[:, :, 1] = -img
unary += 4
fig, axarr = plt.subplots(ncols=unary.shape[-1],
figsize=(unary.shape[-1] * PLOT_SIZE, PLOT_SIZE))
draw_unary(axarr, unary)
fig.tight_layout()
fig.savefig(os.path.join(DIR_IMAGES, 'binary_unary.png'))
# edges, edge_weights = get_uniform_smoothness_pw_single_image(img.shape)
smooth = 1 - np.eye(2)
# y = pygco.cut_grid_graph_simple(unary, pw_cost*0, n_iter=-1)
# labels = pygco.cut_grid_graph_simple(unary_new + np.random.
# randn(unary.shape[0], unary.shape[1], unary.shape[2])*0,
# pw_cost*0, n_iter=-1)
labels = gco.cut_grid_graph_simple(unary, smooth, n_iter=-1)
labels_0 = gco.cut_grid_graph_simple(unary, smooth * 0., n_iter=-1)
fig, axarr = plt.subplots(ncols=3, figsize=(3 * PLOT_SIZE, PLOT_SIZE))
axarr[0].set_title('image')
axarr[0].imshow(img, cmap='gray', interpolation='nearest')
axarr[0].contour(annot, colors='r')
axarr[1].set_title('labeling (smooth=1)')
axarr[1].imshow(labels.reshape(*annot.shape), interpolation='nearest')
axarr[1].contour(annot, colors='w')
axarr[2].set_title('labeling (smooth=0)')
axarr[2].imshow(labels_0.reshape(*annot.shape), interpolation='nearest')
axarr[2].contour(annot, colors='w')
fig.tight_layout()
fig.savefig(os.path.join(DIR_IMAGES, 'binary_labels-4conn.png')), plt.close()
labels = gco.cut_grid_graph_simple(unary, smooth, connect=8, n_iter=-1)
labels_0 = gco.cut_grid_graph_simple(unary, smooth * 0., connect=8, n_iter=-1)
fig, axarr = plt.subplots(ncols=3, figsize=(3 * PLOT_SIZE, PLOT_SIZE))
axarr[0].set_title('image')
axarr[0].imshow(img, cmap='gray', interpolation='nearest')
axarr[0].contour(annot, colors='r')
axarr[1].set_title('labeling (smooth=1)')
axarr[1].imshow(labels.reshape(*annot.shape), interpolation='nearest')
axarr[1].contour(annot, colors='w')
axarr[2].set_title('labeling (smooth=0)')
axarr[2].imshow(labels_0.reshape(*annot.shape), interpolation='nearest')
axarr[2].contour(annot, colors='w')
fig.tight_layout()
fig.savefig(os.path.join(DIR_IMAGES, 'binary_labels-8conn.png')), plt.close()