-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathps5.py
218 lines (191 loc) · 11.1 KB
/
ps5.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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
import numpy as np
from ps_5.helper import lucas_kanade, add_flow_over_im, reduce, gaussian_pyramid, expand, laplacian_pyramid, remap, \
hierarchical_lk, hierarchical_laplacian_lk
from ps_hepers.helpers import imread_from_rep, imshow, np_load, imsave, np_save, imfix_scale, stitch_images, \
get_frames_from_video
"""
Problem Set - 5
Problems: https://docs.google.com/document/d/1Bi2_CThMfoLEf4TCMhyFR7cls6t3LAiSz7YYYV89eoE/pub?embedded=true
"""
def p1_a():
a = imread_from_rep('TestSeq/Shift0', grey_scale=True)
b = imread_from_rep('TestSeq/ShiftR2', grey_scale=True)
c = imread_from_rep('TestSeq/ShiftR5U5', grey_scale=True)
d = imread_from_rep('TestSeq/ShiftR10', grey_scale=True)
e = imread_from_rep('TestSeq/ShiftR20', grey_scale=True)
f = imread_from_rep('TestSeq/ShiftR40', grey_scale=True)
# a
flow1 = lucas_kanade(a, b, (15, 15))
a_flow1 = add_flow_over_im(a, flow1)
flow2 = lucas_kanade(a, c, (47, 47))
a_flow2 = add_flow_over_im(a, flow2)
imsave(a_flow1, 'output/ps5-1-a-1.png')
imsave(a_flow2, 'output/ps5-1-a-2.png')
imshow([a_flow1, a_flow2], ['right shift', 'top right shift'], cmap='gray')
# b
flow_ims = [add_flow_over_im(a, lucas_kanade(a, x, (47, 47))) for x in [d, e, f]]
imshow(flow_ims, ['r10', 'r20', 'r40'], shape=(3, 1))
[imsave(im, 'output/ps5-1-b-%s.png' % (i + 1)) for i, im in zip(range(len(flow_ims)), flow_ims)]
"""
Does it still work? Does it fall apart on any of the pairs?
It does fall apart on all of heavily shifted images.
Motion flow can only be estimated when shift in pixels are small. Since, all the lk flow detector does is compare
gradients ix and iy to time gradient, when time gradient fails (i.e, when delta is too high), it falls apart.
"""
def p1_exp():
im_names = ['TestSeq/Shift0', 'TestSeq/ShiftR5U5']
w_start = 5
w_end = 81
window_range = np.arange(w_start, w_end, 2)
exp_name = ','.join([i.replace('/', '-') for i in im_names])
w_range_as_str = 'w%s-%s' % (w_start, w_end)
a = imread_from_rep(im_names[0], grey_scale=True)
b = imread_from_rep(im_names[1], grey_scale=True)
flow_window_list = np_load(len(window_range), 'objects/flow_%s_%s.npy' % (w_range_as_str, exp_name))
if flow_window_list is None or len(flow_window_list) != len(window_range):
flow_window_list = []
for w in window_range:
print('processing images for motion flow with window size %s' % w)
flow1 = add_flow_over_im(a, lucas_kanade(a, b, (w, w)))
flow_window_list.append((w, flow1))
imsave(flow1, 'observations/flow_%s_w_size_%s.png' % (exp_name, w))
np_save(np.asarray(flow_window_list, dtype=object), 'objects/flow_%s_%s.npy' % (w_range_as_str, exp_name))
def update_exp(x, axs, sliders, buttons):
i = np.abs(window_range - sliders[0].val).argmin()
return [0], [flow_window_list[i][1]]
slider_attr = [{'label': 'Window Size', 'valmin': w_start, 'valmax': w_end, 'valstep': 2}]
imshow(flow_window_list[0][1], [' Detected Optical Flow'], slider_attr=slider_attr, slider_callback=[update_exp])
def p2():
im = imread_from_rep('DataSeq1/yos_img_01', extension='.jpg')
g_py = gaussian_pyramid(im, 4)
imshow(g_py)
[imsave(g_py[i], 'output/ps5-2-a-%s.png' % (i + 1)) for i in range(len(g_py))]
l_py = laplacian_pyramid(im, 4)
l_py = [imfix_scale(i) for i in l_py]
imshow(l_py)
[imsave(l_py[i], 'output/ps5-2-b-%s.png' % (i + 1)) for i in range(len(g_py))]
def p3_exp():
dataset_params = [('DataSeq2/%s', '.png', [0, 1, 2]), ('DataSeq1/yos_img_0%s', '.jpg', [1, 2, 3])]
for prm in dataset_params:
imgs = [imread_from_rep(prm[0] % i, extension=prm[1]) for i in prm[2]]
gpy_s = [gaussian_pyramid(imgs[i], 4) for i in range(len(imgs))]
[imshow([add_flow_over_im(a, lucas_kanade(a, b, (21, 21))) for a, b in zip(gpy_s[j], gpy_s[j + 1])],
['level %s' % i for i in range(len(gpy_s[j]))]) for j in range(len(imgs) - 1)]
def p3():
dataset_params = [('DataSeq1/yos_img_0%s', '.jpg', [1, 2, 3], 1, True),
('DataSeq2/%s', '.png', [0, 1, 2], 2, False)]
for (ds_path, ext, frame_seq, fit_level, upscale), d in zip(dataset_params, range(len(dataset_params))):
imgs = [imread_from_rep(ds_path % i, extension=ext) for i in frame_seq]
gpy_s = [gaussian_pyramid(img, 4, up_scaled=upscale) for img in imgs]
flows = [lucas_kanade(gpy_s[i][fit_level], gpy_s[i + 1][fit_level], (27, 27)) for i in range(len(imgs) - 1)]
arrows = [add_flow_over_im(gpy_s[i][fit_level], flows[i]) for i in range(len(imgs) - 1)]
remaps = [remap(gpy_s[i + 1][fit_level], -flows[i]) for i in range(len(imgs) - 1)]
diffs = [imfix_scale(gpy_s[i][fit_level].astype(np.float32) - remaps[i].astype(np.float32)) for i in
range(len(imgs) - 1)]
imshow(arrows, ['frame %s' % i for i in range(len(imgs))], sup_title='flow arrows')
imshow(remaps, ['frame %s' % i for i in range(len(imgs))], sup_title='remaps')
imsave(stitch_images(arrows), 'output/ps5-3-a-%s.png' % (2 * d + 1))
imsave(stitch_images(imfix_scale(diffs)), 'output/ps5-3-a-%s.png' % (2 * d + 2))
for i in range(len(remaps)):
imsave(gpy_s[i][fit_level], 'observations/3-%s%s.png' % (d, i))
imsave(remaps[i], 'observations/3-%s%sr.png' % (d, i))
def p4_a():
a = imread_from_rep('TestSeq/Shift0')
b = imread_from_rep('TestSeq/ShiftR2')
c = imread_from_rep('TestSeq/ShiftR5U5')
d = imread_from_rep('TestSeq/ShiftR10')
e = imread_from_rep('TestSeq/ShiftR20')
f = imread_from_rep('TestSeq/ShiftR40')
targets = [b, c, d, e, f]
flows = [hierarchical_lk(a, x) for x in targets]
flow_ims = [add_flow_over_im(a, flow) for flow in flows]
imshow([stitch_images([np.abs(flow[:, :, 0]), np.abs(flow[:, :, 1])], axis=0) for flow in flows],
['r2', 'r5u5', 'r10', 'r20', 'r40'], cmap='gray', sup_title='displacement images')
imshow(flow_ims, ['r2', 'r5u5', 'r10', 'r20', 'r40'], sup_title='flow arrows')
remaps = [remap(a, flow) for flow in flows]
imshow([stitch_images([remap_i, target], axis=0) for (remap_i, target) in zip(remaps, targets[:-1])],
['r2', 'r5u5', 'r10', 'r20', 'r40'], sup_title='remap vs actual')
imshow([imfix_scale(remap_i.astype(np.float32) - target.astype(np.float32)) for (remap_i, target) in
zip(remaps, targets[:-1])], sup_title='differences')
imsave(stitch_images(flow_ims), 'output/ps5-4-a-1.png')
imsave(stitch_images([imfix_scale(remap_i.astype(np.float32) - target.astype(np.float32)) for (remap_i, target) in
zip(remaps, targets[:-1])]), 'output/ps5-4-a-2.png')
def p4_bc():
dataset_params = [
('DataSeq1/yos_img_0%s', '.jpg', [1, 2, 3]),
('DataSeq2/%s', '.png', [0, 1, 2]),
]
for (ds_path, ext, frame_seq), d in zip(dataset_params, range(len(dataset_params))):
imgs = [imread_from_rep(ds_path % i, extension=ext) for i in frame_seq]
f_range = range(len(imgs))
flows = [hierarchical_lk(imgs[i], imgs[i + 1]) for i in f_range[:-1]]
arrows = [add_flow_over_im(imgs[i], flows[i]) for i in f_range[:-1]]
remaps = [remap(imgs[i], flows[i]) for i in f_range[:-1]]
diffs = [imgs[i + 1].astype(np.float32) - remaps[i].astype(np.float32) for i in
f_range[:-1]]
imshow([stitch_images(imfix_scale([np.abs(flow[:, :, 0]), np.abs(flow[:, :, 1])]), axis=0) for flow in flows],
range(len(flows)), sup_title='displacement images')
imshow(arrows, ['frame %s' % i for i in range(len(imgs))], sup_title='flow arrows')
imshow(remaps, ['frame %s' % i for i in range(len(imgs))], sup_title='remaps')
imshow([imfix_scale(diff) for diff in diffs], ['frame %s' % i for i in range(len(imgs))],
sup_title='differences')
imsave(stitch_images(arrows), 'output/ps5-4-%s-1.png' % chr(ord('b') + d))
imsave(stitch_images(imfix_scale(diffs)), 'output/ps5-4-%s-2.png' % chr(ord('b') + d))
for i in f_range[:-1]:
imsave(imgs[i + 1], 'observations/4-%s%s.png' % (d, i))
imsave(remaps[i], 'observations/4-%s%sr.png' % (d, i))
def on_video():
imgs = get_frames_from_video('input/car/car.mp4', f_range=range(20, 40))
f_range = range(len(imgs))
[imsave(arrow, 'output/car/img-%s.png' % i) for arrow, i in zip(imgs, f_range)]
flows = [hierarchical_lk(imgs[i], imgs[i + 1]) for i in f_range[:-1]]
[imsave(imfix_scale(np.concatenate([flow, 0 * flow[:, :, 0][:, :, np.newaxis]], axis=2)),
'output/car/flow-%s.png' % i) for flow, i in zip(flows, f_range[:-1])]
arrows = [add_flow_over_im(imgs[i], flows[i] / 200.0, gap=15) for i in f_range[:-1]]
[imsave(arrow, 'output/car/%s.png' % i) for arrow, i in zip(arrows, f_range[:-1])]
def p5():
dataset_params = [('Juggle/%s', '.png', [0, 1, 2])]
for (ds_path, ext, frame_seq), d in zip(dataset_params, range(len(dataset_params))):
imgs = [imread_from_rep(ds_path % i, extension=ext) for i in frame_seq]
f_range = range(len(imgs))
flows = [hierarchical_laplacian_lk(imgs[i], imgs[i + 1]) for i in f_range[:-1]]
arrows = [add_flow_over_im(imgs[i], flows[i]) for i in f_range[:-1]]
remaps = [remap(imgs[i], flows[i]) for i in f_range[:-1]]
diffs = [imgs[i + 1].astype(np.float32) - remaps[i].astype(np.float32) for i in
f_range[:-1]]
imshow([stitch_images(([(flow[:, :, 0]), (flow[:, :, 1])]), axis=0) for flow in flows],
list(range(len(flows))), sup_title='displacement images')
imshow(arrows, ['frame %s' % i for i in range(len(imgs))], sup_title='flow arrows')
imshow(remaps, ['frame %s' % i for i in range(len(imgs))], sup_title='remaps')
imshow([imfix_scale(diff) for diff in diffs], ['frame %s' % i for i in range(len(imgs))],
sup_title='differences')
imsave(stitch_images(arrows), 'output/ps5-5-%s-1.png' % chr(ord('a') + d))
imsave(stitch_images(imfix_scale(diffs)), 'output/ps5-5-%s-2.png' % chr(ord('a') + d))
"""
With some changes I was able to derive motion flow for this sequence. It was hard because the shift in pixels
is too high for lk detection (~35 pixels).
The detected optical flow was not too bad, as it shows the approximate magnitude and direction of shift.
Notice that there was some flow detected over the static pixels around the ball. This is because, the actual
flow was only detected in highest pyramid level and since the image is scaled multiple fold, the pixels around a
moving object are also affected, not to mention the noise added by missing/new pixels at boundary of moving
object.
"""
if __name__ == '__main__':
print('Running p1_a')
p1_a()
print('Running p1_exp')
p1_exp()
print('Running p2')
p2()
print('Running p3_exp')
p3_exp()
print('Running p3')
p3()
print('Running p4_a')
p4_a()
print('Running p4_bc')
p4_bc()
print('Running on_video')
on_video()
print('Running p5')
p5()