import Image
import numpy
import pylab
import vlfeat
from vlfeat.plotop.vl_plotframe import vl_plotframe
if __name__ == "__main__":
""" VL_DEMO_SIFT_EDGE Demo: SIFT: edge treshold
"""
I = numpy.zeros([100, 500])
for i in 10 * (1 + numpy.arange(9)):
d = numpy.round(i / 3.0)
I[50 - d - 1 : 50 + d - 1, i * 5 - 1] = 1
I = numpy.array(I, "f", order="F")
I = 2 * numpy.pi * 8 ** 2 * vlfeat.vl_imsmooth(I, 8)
I = 255 * I
I = numpy.array(I, "f", order="F")
print "sift_edge_0"
ter = [3.5, 5, 7.5, 10]
for te in ter:
f, d = vlfeat.vl_sift(I, peak_thresh=0.0, edge_thresh=te)
pylab.figure()
pylab.gray()
pylab.imshow(I)
vl_plotframe(f, color="k", linewidth=3)
vl_plotframe(f, color="y", linewidth=2)
def vl_phow(im,
verbose=True,
fast=True,
sizes=[4, 6, 8, 10],
step=2,
color='rgb',
floatdescriptors=False,
magnif=6,
windowsize=1.5,
contrastthreshold=0.005):
opts = Options(verbose, fast, sizes, step, color, floatdescriptors, magnif,
windowsize, contrastthreshold)
dsiftOpts = DSiftOptions(opts)
# make sure image is float, otherwise segfault
im = array(im, 'float32')
# Extract the features
imageSize = shape(im)
if im.ndim == 3:
if imageSize[2] != 3:
# "IndexError: tuple index out of range" if both if's are checked at the same time
raise ValueError("Image data in unknown format/shape")
if opts.color == 'gray':
numChannels = 1
if (im.ndim == 2):
im = vl_rgb2gray(im)
else:
numChannels = 3
if (im.ndim == 2):
im = dstack([im, im, im])
if opts.color == 'rgb':
pass
elif opts.color == 'opponent':
# from https://github.com/vlfeat/vlfeat/blob/master/toolbox/sift/vl_phow.m
# Note that the mean differs from the standard definition of opponent
# space and is the regular intesity (for compatibility with
# the contrast thresholding).
# Note also that the mean is added pack to the other two
# components with a small multipliers for monochromatic
# regions.
mu = 0.3 * im[:, :, 0] + 0.59 * im[:, :, 1] + 0.11 * im[:, :, 2]
alpha = 0.01
im = dstack([
mu, (im[:, :, 0] - im[:, :, 1]) / sqrt(2) + alpha * mu,
(im[:, :, 0] + im[:, :, 1] - 2 * im[:, :, 2]) / sqrt(6) +
alpha * mu
])
else:
raise ValueError('Color option ' + str(opts.color) +
' not recognized')
if opts.verbose:
print('{0}: color space: {1}'.format('vl_phow', opts.color))
print('{0}: image size: {1} x {2}'.format('vl_phow', imageSize[0],
imageSize[1]))
print('{0}: sizes: [{1}]'.format('vl_phow', opts.sizes))
frames_all = []
descrs_all = []
for size_of_spatial_bins in opts.sizes:
# from https://github.com/vlfeat/vlfeat/blob/master/toolbox/sift/vl_phow.m
# Recall from VL_DSIFT() that the first descriptor for scale SIZE has
# center located at XC = XMIN + 3/2 SIZE (the Y coordinate is
# similar). It is convenient to align the descriptors at different
# scales so that they have the same geometric centers. For the
# maximum size we pick XMIN = 1 and we get centers starting from
# XC = 1 + 3/2 MAX(OPTS.SIZES). For any other scale we pick XMIN so
# that XMIN + 3/2 SIZE = 1 + 3/2 MAX(OPTS.SIZES).
# In pracrice, the offset must be integer ('bounds'), so the
# alignment works properly only if all OPTS.SZES are even or odd.
off = floor(3.0 / 2 * (max(opts.sizes) - size_of_spatial_bins)) + 1
# smooth the image to the appropriate scale based on the size
# of the SIFT bins
sigma = size_of_spatial_bins / float(opts.magnif)
ims = vl_imsmooth(im, sigma)
# extract dense SIFT features from all channels
frames = []
descrs = []
for k in range(numChannels):
size_of_spatial_bins = int(size_of_spatial_bins)
# vl_dsift does not accept numpy.int64 or similar
f_temp, d_temp = vl_dsift(data=ims[:, :, k],
step=dsiftOpts.step,
size=size_of_spatial_bins,
fast=dsiftOpts.fast,
verbose=dsiftOpts.verbose,
norm=dsiftOpts.norm,
bounds=[off, off, maxint, maxint])
frames.append(f_temp)
descrs.append(d_temp)
frames = array(frames)
descrs = array(descrs)
d_new_shape = [descrs.shape[0] * descrs.shape[1], descrs.shape[2]]
descrs = descrs.reshape(d_new_shape)
# remove low contrast descriptors
# note that for color descriptors the V component is
# thresholded
if (opts.color == 'gray') | (opts.color == 'opponent'):
contrast = frames[0][2, :]
elif opts.color == 'rgb':
contrast = mean(
[frames[0][2, :], frames[1][2, :], frames[2][2, :]], 0)
else:
raise ValueError('Color option ' + str(opts.color) +
' not recognized')
descrs[:, contrast < opts.contrastthreshold] = 0
# save only x,y, and the scale
frames_temp = array(frames[0][0:3, :])
padding = array(size_of_spatial_bins * ones(frames[0][0].shape))
frames_all.append(vstack([frames_temp, padding]))
descrs_all.append(array(descrs))
frames_all = hstack(frames_all)
descrs_all = hstack(descrs_all)
return frames_all, descrs_all
if __name__ == '__main__': """ VL_DEMO_SIFT_PEAK Demo: SIFT: peak treshold """ tmp = uniform(0, 1, (100, 500)) I = numpy.zeros([100, 500]) I.ravel()[pylab.find(tmp.ravel()<=0.005)] = 1 I = (numpy.ones([100,1]) * numpy.r_[0:1:500j]) * I I[:, 0] = 0 I[:, -1] = 0 I[0, :] = 0 I[-1, :] = 0 I = numpy.array(I, 'f', order='F') I = 2 * numpy.pi * 4**2 * vlfeat.vl_imsmooth(I, 4) I = 255 * I print 'sift_peak_0' I = numpy.array(I, 'f', order='F') tpr = [0, 10, 20, 30] for tp in tpr: f, d = vlfeat.vl_sift(I, peak_thresh=tp, edge_thresh=10000) pylab.figure() pylab.gray() pylab.imshow(I) vl_plotframe(f, color='k', linewidth=3) vl_plotframe(f, color='y', linewidth=2)
def dsift(im,
verbose=True,
fast=True,
sizes=[4, 6, 8, 10],
step=2,
color='rgb',
floatdescriptors=False,
magnif=6,
windowsize=1.5,
contrastthreshold=0.005):
opts = Options(verbose, fast, sizes, step, color, floatdescriptors,
magnif, windowsize, contrastthreshold)
dsiftOpts = DSiftOptions(opts)
# make sure image is float, otherwise segfault
im = array(im, 'float32')
# Extract the features
imageSize = shape(im)
if im.ndim == 3:
if imageSize[2] != 3:
# "IndexError: tuple index out of range" if both if's are checked at the same time
raise ValueError("Image data in unknown format/shape")
if opts.color == 'gray':
numChannels = 1
if (im.ndim == 2):
im = vl_rgb2gray(im)
else:
numChannels = 3
if (im.ndim == 2):
im = dstack([im, im, im])
if opts.color == 'rgb':
pass
elif opts.color == 'opponent':
# from https://github.com/vlfeat/vlfeat/blob/master/toolbox/sift/vl_phow.m
# Note that the mean differs from the standard definition of opponent
# space and is the regular intesity (for compatibility with
# the contrast thresholding).
# Note also that the mean is added pack to the other two
# components with a small multipliers for monochromatic
# regions.
mu = 0.3 * im[:, :, 0] + 0.59 * im[:, :, 1] + 0.11 * im[:, :, 2]
alpha = 0.01
im = dstack([mu,
(im[:, :, 0] - im[:, :, 1]) / sqrt(2) + alpha * mu,
(im[:, :, 0] + im[:, :, 1] - 2 * im[:, :, 2]) / sqrt(6) + alpha * mu])
else:
raise ValueError('Color option ' + str(opts.color) + ' not recognized')
if opts.verbose:
great='great'
#print('{0}: color space: {1}'.format('vl_phow', opts.color))
#print('{0}: image size: {1} x {2}'.format('vl_phow', imageSize[0], imageSize[1]))
#print('{0}: sizes: [{1}]'.format('vl_phow', opts.sizes))
frames_all = []
descrs_all = []
for size_of_spatial_bins in opts.sizes:
# from https://github.com/vlfeat/vlfeat/blob/master/toolbox/sift/vl_phow.m
# Recall from VL_DSIFT() that the first descriptor for scale SIZE has
# center located at XC = XMIN + 3/2 SIZE (the Y coordinate is
# similar). It is convenient to align the descriptors at different
# scales so that they have the same geometric centers. For the
# maximum size we pick XMIN = 1 and we get centers starting from
# XC = 1 + 3/2 MAX(OPTS.SIZES). For any other scale we pick XMIN so
# that XMIN + 3/2 SIZE = 1 + 3/2 MAX(OPTS.SIZES).
# In pracrice, the offset must be integer ('bounds'), so the
# alignment works properly only if all OPTS.SZES are even or odd.
off = floor(3.0 / 2 * (max(opts.sizes) - size_of_spatial_bins)) + 1
# smooth the image to the appropriate scale based on the size
# of the SIFT bins
sigma = size_of_spatial_bins / float(opts.magnif)
ims = vl_imsmooth(im, sigma)
# extract dense SIFT features from all channels
frames = []
descrs = []
for k in range(numChannels):
size_of_spatial_bins = int(size_of_spatial_bins)
# vl_dsift does not accept numpy.int64 or similar
f_temp, d_temp = vl_dsift(data=ims[:, :, k],
step=dsiftOpts.step,
size=size_of_spatial_bins,
fast=dsiftOpts.fast,
verbose=dsiftOpts.verbose,
norm=dsiftOpts.norm,
bounds=[off, off, maxint, maxint])
frames.append(f_temp)
descrs.append(d_temp)
frames = array(frames)
descrs = array(descrs)
d_new_shape = [descrs.shape[0] * descrs.shape[1], descrs.shape[2]]
descrs = descrs.reshape(d_new_shape)
# remove low contrast descriptors
# note that for color descriptors the V component is
# thresholded
if (opts.color == 'gray') | (opts.color == 'opponent'):
contrast = frames[0][2, :]
elif opts.color == 'rgb':
contrast = mean([frames[0][2, :], frames[1][2, :], frames[2][2, :]], 0)
else:
raise ValueError('Color option ' + str(opts.color) + ' not recognized')
descrs[:, contrast < opts.contrastthreshold] = 0
# save only x,y, and the scale
frames_temp = array(frames[0][0:3, :])
padding = array(size_of_spatial_bins * ones(frames[0][0].shape))
frames_all.append(vstack([frames_temp, padding]))
descrs_all.append(array(descrs))
frames_all = hstack(frames_all)
descrs_all = hstack(descrs_all)
return frames_all, descrs_all
def test_all(self):
img = numpy.array(Image.open('roofs1.jpg'))
# Test rgb2gray
img_gray = rgb2gray(img)
self.assertEqual(tuple(img_gray.shape), (478, 640))
self.assertTrue(
numpy.allclose(img_gray[:4, :4],
numpy.array([[0.8973, 0.8973, 0.8973, 0.9052],
[0.8973, 0.8973, 0.8973, 0.9052],
[0.8973, 0.8973, 0.9021, 0.9061],
[0.9013, 0.9013, 0.9061, 0.9100]]),
atol=1e-4))
if os.path.exists("img_gray.txt"):
expected = numpy.loadtxt("img_gray.txt", delimiter='\t')
self.assertTrue(numpy.allclose(img_gray, expected, atol=1e-4))
# Test vl_imsmooth
binsize = 8
magnif = 3
img_smooth = vlfeat.vl_imsmooth(img_gray,
math.sqrt((binsize / magnif)**2 -
0.25),
verbose=True)
self.assertEqual(tuple(img_smooth.shape), (478, 640))
self.assertTrue(
numpy.allclose(img_smooth[:4, :4],
numpy.array([[0.8998, 0.9013, 0.9034, 0.9057],
[0.9000, 0.9015, 0.9035, 0.9057],
[0.9002, 0.9017, 0.9036, 0.9057],
[0.9005, 0.9020, 0.9038, 0.9058]]),
atol=1e-4))
if os.path.exists("img_smooth.txt"):
expected = numpy.loadtxt("img_smooth.txt", delimiter='\t')
self.assertTrue(numpy.allclose(img_smooth, expected, atol=1e-4))
# Test vl_dsift
frames, descrs = vlfeat.vl_dsift(img_smooth,
size=binsize,
verbose=True)
frames = numpy.add(frames.transpose(), 1)
descrs = descrs.transpose()
self.assertEqual(tuple(frames.shape), (2, 279664))
self.assertEqual(tuple(descrs.shape), (128, 279664))
self.assertTrue(
numpy.allclose(frames[:, :4],
numpy.array([[13, 13, 13, 13], [13, 14, 15, 16]])))
self.assertTrue(
numpy.allclose(
descrs[:, 0],
numpy.array([
134, 35, 0, 0, 0, 0, 0, 5, 109, 9, 1, 0, 0, 0, 0, 61, 7, 2,
32, 21, 0, 0, 1, 28, 2, 13, 111, 9, 0, 0, 0, 2, 33, 134,
131, 0, 0, 0, 0, 0, 30, 92, 134, 0, 0, 0, 0, 19, 11, 42,
134, 0, 0, 0, 1, 31, 6, 20, 124, 3, 0, 0, 0, 7, 5, 134,
134, 0, 0, 0, 0, 0, 1, 94, 134, 0, 0, 0, 0, 0, 0, 34, 134,
1, 0, 0, 0, 0, 0, 4, 134, 13, 0, 2, 2, 0, 27, 53, 15, 0, 0,
0, 0, 1, 11, 48, 27, 2, 0, 0, 0, 0, 0, 5, 28, 16, 1, 0, 0,
0, 0, 2, 13, 16, 4, 5, 4, 0
])))
if os.path.exists("dsift_frames.txt"):
expected = numpy.loadtxt("dsift_frames.txt", delimiter='\t')
self.assertTrue(numpy.allclose(frames, expected))
if os.path.exists("dsift_descrs.txt"):
expected = numpy.loadtxt("dsift_descrs.txt", delimiter='\t')
self.assertTrue(
numpy.linalg.norm(expected - descrs) < 28) # rounding errors?
# Test vl_kmeans
centers, assigns = vlfeat.vl_kmeans(numpy.array(
[[1], [2], [3], [10], [11], [12]], dtype='f'),
2,
ret_quantize=True,
verbose=True)
self.assertTrue(numpy.allclose(centers, numpy.array([[2], [11]])))
self.assertTrue(
numpy.allclose(assigns, numpy.array([0, 0, 0, 1, 1, 1])))
centers, assigns = vlfeat.vl_kmeans(numpy.array(
[[1, 0], [2, 0], [3, 0], [10, 1], [11, 1], [12, 1]], dtype='f'),
2,
ret_quantize=True)
self.assertTrue(numpy.allclose(centers,
numpy.array([[11, 1],
[2,
0]]))) # order swapped?
self.assertTrue(
numpy.allclose(assigns, numpy.array([1, 1, 1, 0, 0, 0])))
# Test vl_gmm
if os.path.exists("gmm_data.txt"):
data = numpy.loadtxt("gmm_data.txt", delimiter='\t').transpose()
else:
data = numpy.random.rand(5000, 2)
means, covariances, priors, ll, posteriors = vlfeat.vl_gmm(
data, 30, verbose=True, ret_loglikelihood=True, ret_posterior=True)
self.assertEqual(tuple(means.shape), (30, 2))
self.assertEqual(tuple(covariances.shape), (30, 2))
self.assertEqual(tuple(priors.shape), (30, ))
self.assertEqual(tuple(posteriors.shape), (5000, 30))
if os.path.exists("gmm_means.txt"):
expected = numpy.loadtxt("gmm_means.txt",
delimiter='\t').transpose()
self.assertTrue(numpy.allclose(means, expected, atol=1e-4))
if os.path.exists("gmm_covariances.txt"):
expected = numpy.loadtxt("gmm_covariances.txt",
delimiter='\t').transpose()
self.assertTrue(numpy.allclose(covariances, expected, atol=1e-4))
if os.path.exists("gmm_priors.txt"):
expected = numpy.loadtxt("gmm_priors.txt",
delimiter='\t').transpose()
self.assertTrue(numpy.allclose(priors, expected, atol=1e-4))
if os.path.exists("gmm_posteriors.txt"):
expected = numpy.loadtxt("gmm_posteriors.txt",
delimiter='\t').transpose()
self.assertTrue(numpy.allclose(posteriors, expected, atol=1e-3))
# Test vl_fisher
if os.path.exists("fisher_data.txt"):
data = numpy.loadtxt("fisher_data.txt", delimiter='\t').transpose()
else:
data = numpy.random.rand(1000, 2)
encoding = vlfeat.vl_fisher(data,
means,
covariances,
priors,
verbose=True)
self.assertEqual(tuple(encoding.shape), (120, ))
if os.path.exists("fisher_encoding.txt"):
expected = numpy.loadtxt("fisher_encoding.txt",
delimiter='\t').transpose()
self.assertTrue(numpy.allclose(encoding, expected, atol=1e-4))