def test_sift_non_float_descriptors():
i = half_img.copy()
frames, descriptors = sift(i, compute_descriptor=True)
assert_allclose(frames[0], [2.16217, 128.056, 2.13029, 5.9325], rtol=1e-3)
assert_allclose(descriptors[0, 84:88], [4, 14, 32, 137])
assert frames.shape[0] == 358
def test_sift_non_float_descriptors():
i = half_img.copy()
frames, descriptors = sift(i, compute_descriptor=True)
assert_allclose(frames[0], [2.16217, 128.056, 2.13029, -4.3617], rtol=1e-3)
assert_allclose(descriptors[0, 47:52], [53, 131, 137, 32, 14])
assert frames.shape[0] == 358
def test_sift_non_float_descriptors():
i = half_img.copy()
frames, descriptors = sift(i, compute_descriptor=True)
assert_allclose(frames[0], [2.16217, 128.056, 2.13029, -4.3617], rtol=1e-3)
assert_allclose(descriptors[0, 47:52], [53, 131, 137, 32, 14])
assert frames.shape[0] == 358
def test_dsift_sift(window_size):
bin_size = 4
magnif = 3
scale = bin_size / magnif
img_smooth = gaussian(img, sigma=sqrt(scale**2 - 0.25))
f, d = dsift(img_smooth,
size=bin_size,
step=10,
window_size=window_size,
float_descriptors=True)
num_keypoints = f.shape[0]
f_ = np.column_stack([
f,
np.ones(shape=(num_keypoints, )) * scale,
np.zeros(shape=(num_keypoints, ))
])
f_, d_ = sift(img,
magnification=magnif,
frames=f_,
first_octave=-1,
n_levels=5,
compute_descriptor=True,
float_descriptors=True,
window_size=window_size)
err = np.std(d - d_) / np.std(d)
assert err < 0.1
def test_sift_shape(compute_descriptor):
result = sift(img, compute_descriptor=compute_descriptor)
if compute_descriptor:
assert result[0].shape[1] == 4
assert result[1].shape[1] == 128
assert result[0].shape[0] == result[1].shape[0]
else:
assert result.shape[1] == 4
def test_sift_sort_user_defined_scales():
i = half_img.copy()
frames = np.array([[4., 5., 3., 0], [3., 2., 2., np.pi / 2],
[8., 6., 1., 1.]])
new_frames, descriptors = sift(i, frames=frames, compute_descriptor=True)
assert_allclose(new_frames[0], frames[-1], rtol=1e-3)
assert_allclose(descriptors[0, -5:], [46, 14, 0, 0, 0])
assert new_frames.shape[0] == 3
def test_sift_force_orientations():
i = half_img.copy()
frames = np.array([[4., 5., 2., 0], [3., 2., 3., np.pi / 2],
[8., 6., 4., 1.]])
new_frames, descriptors = sift(i, frames=frames, compute_descriptor=True,
force_orientations=True)
assert_allclose(new_frames[0], [4, 5, 2, 4.6239], rtol=1e-3)
assert_allclose(descriptors[0, :5], [5, 1, 0, 0, 1])
assert new_frames.shape[0] == 3
def test_sift_force_orientations():
frames = np.array([[6.90227, 242.036, 1.84193, 2.63054],
[5.28098, 342.099, 1.90079, 5.07138],
[22.1369, 273.583, 2.06255, 1.56508]])
new_frames, descriptors = sift(img,
frames=frames,
compute_descriptor=True,
force_orientations=True)
assert new_frames.shape[0] == 4
assert_allclose(new_frames[0], [6.90227, 242.036, 1.84193, 2.63051],
rtol=1e-3)
assert_allclose(descriptors[0, :10],
[107, 171, 0, 0, 0, 0, 0, 11, 144, 171])
def test_sift_sort_user_defined_scales():
frames = np.array([[5.28098, 342.099, 1.90079, 5.07138],
[6.90227, 242.036, 1.84193, 2.63054],
[22.1369, 273.583, 2.06255, 1.56508]])
new_frames, descriptors = sift(img, frames=frames, compute_descriptor=True)
assert new_frames.shape[0] == 3
assert_allclose(new_frames[0], frames[1], rtol=1e-3)
assert_allclose(new_frames[1], frames[0], rtol=1e-3)
assert_allclose(new_frames[2], frames[2], rtol=1e-3)
assert_allclose(descriptors[0, :10],
[107, 171, 0, 0, 0, 0, 0, 11, 144, 171])
assert_allclose(descriptors[1, :10], [17, 0, 0, 11, 20, 1, 16, 73, 95, 11])
assert_allclose(descriptors[2, :10], [134, 15, 0, 0, 1, 1, 0, 0, 134, 73])
def test_sift_detector():
f = sift(img, first_octave=-1, peak_thresh=0.01)
frames = actual_frames.copy()
# scale the components so that 1 pixel erro in x,y,z is equal to a
# 10-th of angle
scale = (20 / np.pi)
frames[:, -1] = np.mod(frames[:, -1], np.pi * 2) * scale
f[:, -1] = np.mod(f[:, -1], np.pi * 2) * scale
D2 = cdist(frames, f)
d2 = D2.min(axis=1)
d2.sort()
err = np.sqrt(d2)
quant80 = round(0.8 * f.shape[0])
# check for less than one pixel error at 80% quantile
assert err[quant80] < 1
def test_sift_user_defined_frames():
frames, descriptors = sift(img,
frames=actual_frames,
first_octave=-1,
compute_descriptor=True,
float_descriptors=True,
norm_thresh=0,
verbose=False)
D2 = cdist(actual_frames, frames)
index = D2.argmin(axis=1)
descriptors = descriptors[index, :]
t1 = np.mean(np.sqrt(np.sum((actual_descriptors - descriptors)**2,
axis=1)))
t2 = np.mean(np.sqrt(np.sum(actual_descriptors**2, axis=1)))
err = t1 / t2
assert err < 0.1
def test_sift_n_frames():
i = img.copy()
frames = sift(i)
assert_allclose(frames[0], [2.16217, 128.056, 2.13029, 5.9325], rtol=1e-3)
assert frames.shape[0] == 730
def test_sift_non_float_descriptors():
frames, descriptors = sift(img, compute_descriptor=True)
assert_allclose(frames[0], [5.28098, 342.099, 1.90079, 5.07138], rtol=1e-3)
assert_allclose(descriptors[0, :5], [17, 0, 0, 11, 20])
assert frames.shape[0] == 797
