def shape_index_hist(img, scale, n_bins, tonal_scale):
si, si_c = ipcv.shape_index(img, scale, orientations=False)
# sp.misc.imsave('woop%.2f.png' % scale, si)
limits = (-np.pi / 2, np.pi / 2)
si_iso = ipcv.misc.isophotes(si, n_bins, limits, tonal_scale)
hist = np.sum(si_iso * si_c, axis=(1, 2))
hist /= np.sum(hist)
return hist
def shape_index_hist(img, scale, n_bins, tonal_scale):
si, si_c = ipcv.shape_index(img, scale, orientations=False)
# sp.misc.imsave('woop%.2f.png' % scale, si)
limits = (-np.pi/2, np.pi/2)
si_iso = ipcv.misc.isophotes(si, n_bins, limits, tonal_scale)
hist = np.sum(si_iso * si_c, axis=(1,2))
hist /= np.sum(hist)
return hist
def transform(self, img):
hists_shape = (self.n_bins, self.ori_n_bins, len(self.scales))
hists = np.empty(hists_shape)
si, si_c, si_o, si_om = ipcv.shape_index(img,
self.ori_detect_scale,
orientations=True)
# Detect dominant orientation
ori_detect_n_bins = self.ori_n_bins * 2
iso_si_o = ipcv.misc.isophotes(si_o, ori_detect_n_bins,
(-np.pi / 2, np.pi / 2),
self.ori_tonal_scale, 'von_mises')
ori_hist = np.sum(iso_si_o, axis=(1, 2))
ori_hist /= np.sum(ori_hist)
# print(ori_hist)
max_idx = np.argmax(ori_hist)
ori_offset = max_idx / float(ori_detect_n_bins) * np.pi - np.pi / 2
# print(max_idx, max_idx/float(ori_detect_n_bins), ori_offset)
for s_idx, s in enumerate(self.scales):
si, si_c, si_o, si_om = ipcv.shape_index(img, s, orientations=True)
# Shift according to dominant orientation
si_o = np.mod(si_o - ori_offset + np.pi / 2, np.pi) - np.pi / 2
# Smooth bin contributions (= soft isophote images)
iso_si = ipcv.misc.isophotes(si, self.n_bins,
(-np.pi / 2, np.pi / 2),
self.tonal_scale)
iso_si_o = ipcv.misc.isophotes(si_o, self.ori_n_bins,
(-np.pi / 2, np.pi / 2),
self.ori_tonal_scale, 'von_mises')
# Bin contributions for the joint histogram
iso_j = (iso_si[:, np.newaxis, ...] * si_c *
iso_si_o[np.newaxis, ...] * si_om)
# Summarize bin contributions in the joint histograms
hists[:, :, s_idx] = np.sum(iso_j, axis=(2, 3))
if not self.joint_hist:
hists_si = np.sum(hists, axis=1)
hists_ori = np.sum(hists, axis=0)
hists_si = ipcv.misc.normalize(hists_si, self.norm)
hists_ori = ipcv.misc.normalize(hists_ori, self.norm)
hists = np.hstack([np.ravel(hists_si), np.ravel(hists_ori)])
else:
hists = ipcv.misc.normalize(hists, self.norm)
if self.flat:
hists = np.ravel(hists)
return hists
def visualize_si():
''' Visualize the shape index responses. '''
img = sp.misc.imread('data/patterns.png', flatten=True)
si, si_c, si_o, si_om = shape_index(img, 2.5, orientations=True, fft=True)
imsave('si/patterns-si.png', si)
imsave('si/patterns-si_c.png', si_c)
imsave('si/patterns-si_weighted.png', si * si_c)
imsave('si/patterns-si_o.png', si_o)
imsave('si/patterns-si_om.png', si_om)
imsave('si/patterns-si_o_weighted.png', si_o * si_om)
def visualize_si():
''' Visualize the shape index responses. '''
img = sp.misc.imread('data/patterns.png', flatten=True)
si, si_c, si_o, si_om = shape_index(img, 2.5, orientations=True, fft=True)
imsave('si/patterns-si.png', si)
imsave('si/patterns-si_c.png', si_c)
imsave('si/patterns-si_weighted.png', si*si_c)
imsave('si/patterns-si_o.png', si_o)
imsave('si/patterns-si_om.png', si_om)
imsave('si/patterns-si_o_weighted.png', si_o*si_om)
def transform(self, img):
hists_shape = (self.n_bins, self.ori_n_bins, len(self.scales))
hists = np.empty(hists_shape)
si, si_c, si_o, si_om = ipcv.shape_index(img, self.ori_detect_scale, orientations=True)
# Detect dominant orientation
ori_detect_n_bins = self.ori_n_bins*2
iso_si_o = ipcv.misc.isophotes(si_o, ori_detect_n_bins, (-np.pi/2, np.pi/2),
self.ori_tonal_scale, 'von_mises')
ori_hist = np.sum(iso_si_o, axis=(1,2))
ori_hist /= np.sum(ori_hist)
# print(ori_hist)
max_idx = np.argmax(ori_hist)
ori_offset = max_idx/float(ori_detect_n_bins)*np.pi-np.pi/2
# print(max_idx, max_idx/float(ori_detect_n_bins), ori_offset)
for s_idx, s in enumerate(self.scales):
si, si_c, si_o, si_om = ipcv.shape_index(img, s, orientations=True)
# Shift according to dominant orientation
si_o = np.mod(si_o-ori_offset+np.pi/2, np.pi)-np.pi/2
# Smooth bin contributions (= soft isophote images)
iso_si = ipcv.misc.isophotes(si, self.n_bins, (-np.pi/2, np.pi/2),
self.tonal_scale)
iso_si_o = ipcv.misc.isophotes(si_o, self.ori_n_bins, (-np.pi/2, np.pi/2),
self.ori_tonal_scale, 'von_mises')
# Bin contributions for the joint histogram
iso_j = (iso_si[:, np.newaxis, ...] * si_c
* iso_si_o[np.newaxis, ...] * si_om)
# Summarize bin contributions in the joint histograms
hists[:, :, s_idx] = np.sum(iso_j, axis=(2, 3))
if not self.joint_hist:
hists_si = np.sum(hists, axis=1)
hists_ori = np.sum(hists, axis=0)
hists_si = ipcv.misc.normalize(hists_si, self.norm)
hists_ori = ipcv.misc.normalize(hists_ori, self.norm)
hists = np.hstack([np.ravel(hists_si), np.ravel(hists_ori)])
else:
hists = ipcv.misc.normalize(hists, self.norm)
if self.flat:
hists = np.ravel(hists)
return hists
def multiscale_shape_index():
dataset = data.UIUCTex()
imgs = memory.cache(dataset.imgs)()
img_idx = 401
img = imgs[img_idx]
dirpath = os.path.join('multiscale_shape_index', dataset.name, 'img%.3d'%img_idx)
img_save(img, os.path.join(dirpath, 'img.png'))
scales = [1, 2, 3, 4, 5, 6, 7, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 32, 34, 36, 38]
for scale in scales:
si, si_c, si_o, si_om = ipcv.shape_index(img, scale, orientations=True)
img_save(si, os.path.join(dirpath, 'si_scale%.2d.png' % scale))
img_save(si_c, os.path.join(dirpath, 'si-c_scale%.2d.png' % scale))
img_save(si_o, os.path.join(dirpath, 'si-o_scale%.2d.png' % scale))
img_save(si_om, os.path.join(dirpath, 'si-om_scale%.2d.png' % scale))
def multiscale_shape_index():
dataset = data.UIUCTex()
imgs = memory.cache(dataset.imgs)()
img_idx = 401
img = imgs[img_idx]
dirpath = os.path.join('multiscale_shape_index', dataset.name,
'img%.3d' % img_idx)
img_save(img, os.path.join(dirpath, 'img.png'))
scales = [
1, 2, 3, 4, 5, 6, 7, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 32,
34, 36, 38
]
for scale in scales:
si, si_c, si_o, si_om = ipcv.shape_index(img, scale, orientations=True)
img_save(si, os.path.join(dirpath, 'si_scale%.2d.png' % scale))
img_save(si_c, os.path.join(dirpath, 'si-c_scale%.2d.png' % scale))
img_save(si_o, os.path.join(dirpath, 'si-o_scale%.2d.png' % scale))
img_save(si_om, os.path.join(dirpath, 'si-om_scale%.2d.png' % scale))
def visualize_si_orientation():
''' Visualize the shape index orientation. '''
img = sp.misc.imread('data/rings.png', flatten=True)
si, si_c, si_o, si_om = shape_index(img, 2.5, orientations=True, fft=True)
import matplotlib.pyplot as plt
plt.figure()
plt.imshow(img, cmap=plt.gray())
X, Y = np.mgrid[0:si_o.shape[0], 0:si_o.shape[1]]
u = np.cos(si_o) * si_c / np.max(si_c)
v = np.sin(si_o) * si_c / np.max(si_c)
quiver_mask = si_c > 10
step = 3
quiver_mask = quiver_mask[::step, ::step]
Y = (Y[::step, ::step])[quiver_mask]
X = (X[::step, ::step])[quiver_mask]
u = (u[::step, ::step])[quiver_mask]
v = (v[::step, ::step])[quiver_mask]
plt.quiver(Y, X, u, v, pivot='mid', scale=30, color='r')
plt.savefig('si/rings-si_o.pdf')
def visualize_si_fiducial_orientation():
''' Visualize the shape index orientation with a fiducial coordinate
system.'''
img = sp.misc.imread('data/rings.png', flatten=True)
si, si_c, si_o, si_om = shape_index(img, 2.5, orientations=True, fft=True)
# No fiducial coordinate system.
imsave('si/rings-si_o.png', si_o)
imsave('si/rings-si_o_weighted.png', si_o * si_om)
# Let each pixel have its own coordinate system where the origin is
# the image center and the first axis is the vector from the origin to
# the pixel.
h, w = img.shape[:2]
y = np.linspace(-h / 2., h / 2., h)
x = np.linspace(-w / 2., w / 2., w)
xv, yv = np.meshgrid(x, y)
offsets = np.arctan(yv / (xv + 1e-10))
si_o = np.mod(si_o + offsets, np.pi)
imsave('si/rings-si_o_fiducial.png', si_o)
imsave('si/rings-si_o_fiducial_weighted.png', si_o * si_om)
def visualize_si_fiducial_orientation():
''' Visualize the shape index orientation with a fiducial coordinate
system.'''
img = sp.misc.imread('data/rings.png', flatten=True)
si, si_c, si_o, si_om = shape_index(img, 2.5, orientations=True, fft=True)
# No fiducial coordinate system.
imsave('si/rings-si_o.png', si_o)
imsave('si/rings-si_o_weighted.png', si_o*si_om)
# Let each pixel have its own coordinate system where the origin is
# the image center and the first axis is the vector from the origin to
# the pixel.
h, w = img.shape[:2]
y = np.linspace(-h/2., h/2., h)
x = np.linspace(-w/2., w/2., w)
xv, yv = np.meshgrid(x, y)
offsets = np.arctan(yv/(xv+1e-10))
si_o = np.mod(si_o+offsets, np.pi)
imsave('si/rings-si_o_fiducial.png', si_o)
imsave('si/rings-si_o_fiducial_weighted.png', si_o*si_om)
def visualize_si_orientation():
''' Visualize the shape index orientation. '''
img = sp.misc.imread('data/rings.png', flatten=True)
si, si_c, si_o, si_om = shape_index(img, 2.5, orientations=True, fft=True)
import matplotlib.pyplot as plt
plt.figure()
plt.imshow(img, cmap=plt.gray())
X, Y = np.mgrid[0:si_o.shape[0], 0:si_o.shape[1]]
u = np.cos(si_o)*si_c/np.max(si_c)
v = np.sin(si_o)*si_c/np.max(si_c)
quiver_mask = si_c > 10
step = 3
quiver_mask = quiver_mask[::step, ::step]
Y = (Y[::step, ::step])[quiver_mask]
X = (X[::step, ::step])[quiver_mask]
u = (u[::step, ::step])[quiver_mask]
v = (v[::step, ::step])[quiver_mask]
plt.quiver(Y, X, u, v,
pivot='mid', scale=30,
color='r')
plt.savefig('si/rings-si_o.pdf')