-
Notifications
You must be signed in to change notification settings - Fork 1
/
frame_interp.py
executable file
·188 lines (150 loc) · 7.11 KB
/
frame_interp.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
"""
Version: 1.0
Summary: Phase based motipn frame prediction and interpolation function
Author: suxing liu
Author-email: [email protected]
USAGE:
from frame_interp import interpolate_frame
"""
import numpy as np
from pyPyrTools import SCFpyr
from skimage import color
from skimage import transform
def decompose(img, ht, n_orientations, t_width, scale, n_scales, xp=np):
"""decompose image into different pyramids levels"""
lab = xp.array(color.rgb2lab(img)) / 255.
pyramids = {'pyramids': [], 'high_pass': [], 'low_pass': [], 'phase': [], 'amplitude': [], 'pind': 0}
for i in range(img.shape[-1]):
pyr = SCFpyr(lab[..., i], ht, n_orientations - 1, t_width, scale, n_scales, xp)
pyramids['pyramids'].append(pyr)
pyramids['high_pass'].append(pyr.pyrHigh())
pyramids['low_pass'].append(pyr.pyrLow())
pyramids['phase'].append([xp.angle(pyr.pyr[level]) for level in range(1, len(pyr.pyr) - 1)])
pyramids['amplitude'].append([xp.abs(pyr.pyr[level]) for level in range(1, len(pyr.pyr) - 1)])
return pyramids
def compute_phase_difference(L, R, *args):
"""compute phase difference between images"""
xp = L['pyramids'][0].xp
phase_diff_out = []
for i in range(len(L['phase'])):
phase_diff = [xp.arctan2(xp.sin(R['phase'][i][j] - L['phase'][i][j]), xp.cos(R['phase'][i][j] - L['phase'][i][j]))
for j in range(len(L['phase'][i]))]
phase_diff_new = shift_correction(phase_diff, L['pyramids'][i], *args)
unwrapped_phase_diff = []
for j in range(len(phase_diff_new)):
unwrapped_phase_diff.append(unwrap(xp.stack([phase_diff_new[j], list(phase_diff)[j]], 0),
xp=L['pyramids'][0].xp)[0])
phase_diff_out.append(unwrapped_phase_diff)
return phase_diff_out
def shift_correction(pyr, pyramid, *args):
"""correct the shift between pyramid levels"""
n_high_elems = pyramid.pyrSize[0]
n_low_elems = pyramid.pyrSize[-1]
corrected_pyr = list(pyr)
corrected_pyr.insert(0, np.zeros(n_high_elems))
corrected_pyr.append(np.zeros(n_low_elems))
n_levels = pyramid.spyrHt()
n_bands = pyramid.numBands()
for level in range(n_levels - 1, -1, -1):
corrected_level = correct_level(corrected_pyr, pyramid, level, *args)
start_ind = 1 + n_bands * level
corrected_pyr[start_ind:start_ind+n_bands] = corrected_level
corrected_pyr = corrected_pyr[1:len(corrected_pyr) - 1]
return corrected_pyr
def correct_level(pyr, pyramid, level, *args):
"""correct pyramid levels"""
scale = args[0]
limit = args[1]
n_levels = pyramid.spyrHt()
n_bands = pyramid.numBands()
out_level = []
if level < n_levels - 1:
dims = pyramid.pyrSize[1+n_bands*level]
for band in range(n_bands):
index_lo = pyramid.bandIndex(level + 1, band)
low_level_small = pyr[index_lo]
if pyramid.xp.__name__ == 'numpy':
low_level = transform.resize(low_level_small, dims, mode='reflect').astype('float32')
else:
low_level = pyramid.xp.array(transform.resize(pyramid.xp.asnumpy(low_level_small),
dims, mode='reflect').astype('float32'))
index_hi = pyramid.bandIndex(level, band)
high_level = pyr[index_hi]
unwrapped = pyramid.xp.stack([low_level.reshape(-1) / scale, high_level.reshape(-1)], 0)
unwrapped = unwrap(unwrapped, xp=pyramid.xp)
high_level = unwrapped[1]
high_level = pyramid.xp.reshape(high_level, dims)
angle_diff = pyramid.xp.arctan2(pyramid.xp.sin(high_level-low_level/scale),
pyramid.xp.cos(high_level-low_level/scale))
to_fix = pyramid.xp.abs(angle_diff) > (np.pi / 2)
high_level[to_fix] = low_level[to_fix] / scale
if limit > 0:
to_fix = pyramid.xp.abs(high_level) > (limit * np.pi / scale ** (n_levels - level))
high_level[to_fix] = low_level[to_fix] / scale
out_level.append(high_level)
if level == n_levels - 1:
for band in range(n_bands):
index_lo = pyramid.bandIndex(level, band)
low_level = pyr[index_lo]
if limit > 0:
to_fix = pyramid.xp.abs(low_level) > (limit * np.pi / scale ** (n_levels - level))
low_level[to_fix] = 0.
out_level.append(low_level)
return out_level
def unwrap(p, cutoff=np.pi, xp=np):
"""correct pyramid levels"""
def local_unwrap(p, cutoff):
dp = p[1] - p[0]
dps = xp.mod(dp + np.pi, 2 * np.pi) - np.pi
dps[xp.logical_and(dps == -np.pi, dp > 0)] = np.pi
dp_corr = dps - dp
dp_corr[xp.abs(dp) < cutoff] = 0.
p[1] += dp_corr
return p
shape = p.shape
p = xp.reshape(p, (shape[0], np.prod(shape[1:])))
q = local_unwrap(p, cutoff)
q = xp.reshape(q, shape)
return q
def interpolate_pyramid(L, R, phase_diff, alpha):
"""compute the interpolation between pyramid levels"""
new_pyr = []
for i in range(len(phase_diff)):
new_pyr.append([])
high_pass = L['high_pass'][i] if alpha < 0.5 else R['high_pass'][i]
low_pass = (1 - alpha) * L['low_pass'][i] + alpha * R['low_pass'][i]
new_pyr[i].append(high_pass)
for k in range(len(R['phase'][i])):
new_phase = R['phase'][i][k] + (alpha - 1) * phase_diff[i][k]
new_amplitude = (1 - alpha) * L['amplitude'][i][k] + alpha * R['amplitude'][i][k]
mid_band = new_amplitude * np.e ** (1j * new_phase)
new_pyr[i].append(mid_band)
new_pyr[i].append(low_pass)
return new_pyr
def reconstruct_image(pyr):
"""reconstruct images based on pyramid level"""
xp = pyr['pyramids'][0].xp
out_img = xp.zeros((pyr['pyramids'][0].pyrSize[0][0], pyr['pyramids'][0].pyrSize[0][1], 3))
for i, pyr in enumerate(pyr['pyramids']):
out_img[..., i] = pyr.reconPyr('all', 'all')
if xp.__name__ == 'numpy':
out_img = color.lab2rgb(out_img * 255.)
else:
out_img = color.lab2rgb(xp.asnumpy(out_img) * 255.)
return out_img
def interpolate_frame(img1, img2, n_frames=1, n_orientations=8, t_width=1, scale=0.5, limit=.4, min_size=15, max_levels=23, xp=np):
"""compute the interpolated images"""
h, w, l = img1.shape
n_scales = min(np.ceil(np.log2(min((h, w))) / np.log2(1. / scale) -
(np.log2(min_size) / np.log2(1 / scale))).astype('int'), max_levels)
step = 1. / (n_frames + 1)
L = decompose(img1, n_scales, n_orientations, t_width, scale, n_scales, xp)
R = decompose(img2, n_scales, n_orientations, t_width, scale, n_scales, xp)
phase_diff = compute_phase_difference(L, R, scale, limit)
new_frames = []
for j in range(n_frames):
new_pyr = interpolate_pyramid(L, R, phase_diff, step * (j + 1))
for i, pyr in enumerate(L['pyramids']):
pyr.pyr = new_pyr[i]
new_frames.append(reconstruct_image(L))
return new_frames