-
Notifications
You must be signed in to change notification settings - Fork 31
/
neural_scene_graph_manipulation.py
215 lines (166 loc) · 7.09 KB
/
neural_scene_graph_manipulation.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
import tensorflow as tf
import numpy as np
from matplotlib import pyplot as plt
def spin_object(obj_pose, steps=11, min=-np.pi/30, max=np.pi/30):
obj_pose = np.reshape(obj_pose, [-1, 6])
rotation_poses = np.repeat(obj_pose, steps, axis=0)
rotation_poses[:, 3] += np.linspace(min, max, steps)
return np.reshape(rotation_poses, [steps, -1, 3])
def dancing_objects(obj_i, steps=30, min=-np.pi*1/5, max=np.pi*1/5):
obj_dancing = -1 * np.ones([steps, obj_i.shape[0], obj_i.shape[1]])
for k in range(obj_i.shape[0]//2):
obj_to_spin = obj_i[2 * k:2 * (k + 1)]
if obj_to_spin[0, 0] != -1:
spined_obj = spin_object(obj_to_spin, steps, min, max)
obj_dancing[:, 2 * k:2 * (k + 1), :] = spined_obj
return obj_dancing
def move_along_track(objs):
return objs
def ghost_street(objs):
selected_obj = np.random.choice
return objs
def remove_obj_from_set(obj_set, obj_meta, rm_obj):
rm_id = np.where(np.equal(obj_meta[:, 0], rm_obj))[0]
rm_lines = np.where(obj_set[:, :, 1] == rm_id)
obj_set[rm_lines] = np.array([-1., 0., 0.])
obj_set[rm_lines[0], rm_lines[1] - 1] = np.array([-1., 1., -1.])
return obj_set
def manipulate_obj_pose(manipulation, obj, obj_meta, i, rm_obj=None):
''' composing scenes via neural scene graph manipulation functions such as:
- rotate
- dance
- move
- switch_location
- oversample
- traffic_jam
- handcrafted_set_xx
Args:
manipulation: requested manipulation
obj: scene graph nodes and edges at all times in the sequence
obj_meta: object metadata (latent codes, size)
i: frame in sequence
rm_obj: removable objects
Returns:
render_set
'''
render_set = []
if rm_obj is not None:
if isinstance(rm_obj, list):
for obj_instance in rm_obj:
obj = remove_obj_from_set(obj, obj_meta, obj_instance)
else:
obj = remove_obj_from_set(obj, obj_meta, rm_obj)
if manipulation is None:
obj_i = tf.cast(obj[i], tf.float32)
obj_i = tf.reshape(obj_i, [obj_i.shape[0] // 2, 2 * 3])
render_set.append([obj_i])
elif manipulation == 'rotate':
select_obj = np.random.choice(np.linspace(0, len(obj[0])//2 - 1,len(obj[0])//2).astype(np.int32))
rotate_obj = spin_object(obj[i, select_obj*2:(select_obj*2+2)])
obj = np.repeat(obj[i][None], rotate_obj.shape[0], axis=0)
obj = np.ones_like(obj) * -1.
obj[..., 0:2, :] = rotate_obj
for obj_i in obj:
obj_i = tf.cast(obj_i, tf.float32)
obj_i = tf.reshape(obj_i, [obj_i.shape[0] // 2, 2 * 3])
render_set.append([obj_i])
elif manipulation == 'dance':
obj = dancing_objects(obj[i])
for obj_i in obj:
obj_i = tf.cast(obj_i, tf.float32)
obj_i = tf.reshape(obj_i, [obj_i.shape[0] // 2, 2 * 3])
render_set.append([obj_i])
elif manipulation == 'move':
obj_flat = np.reshape(obj, [-1, 6])
ids = np.unique(obj_flat[:, 4])
rand_moving_id = np.random.choice(ids)
render_obj_poses = obj_flat[np.where(obj_flat[..., 4] == rand_moving_id)]
obj = -1 * np.ones([render_obj_poses.shape[0], obj.shape[1]//2, 6])
obj[..., 4] = 0.
obj[:, 0, :] = render_obj_poses
for obj_i in obj:
obj_i = tf.cast(obj_i, tf.float32)
render_set.append([obj_i])
elif manipulation == 'switch_location':
obj_i = np.reshape(obj[i], [-1, 6])
obj_ids = obj_i[:, 4]
np.random.shuffle(obj_ids)
obj_i[:, 4] = obj_ids
obj_i = tf.cast(obj[i], tf.float32)
obj_i = tf.reshape(obj_i, [obj_i.shape[0] // 2, 2 * 3])
render_set.append([obj_i])
elif manipulation == 'oversample':
factor = 2
N_obj = len(obj[0]) // 2 * factor
obj_flat = np.reshape(obj, [-1, 6])
avg_z = np.mean(obj_flat[:, 1])
std_z = np.std(obj_flat[:, 1])
avg_angle = np.mean(obj_flat[:, 3]) + 0.04
std_angle = np.std(obj_flat[:, 3])
max_x = np.abs(obj_flat[:, 0]).max()
max_y = 20 # np.abs(obj_flat[:, 2]).max()
thresh = np.mean(obj_meta[1:, 1]) * 1.2
obj_pose_flat = []
obj_flat = []
for i in range(N_obj):
close = True
j = 0
while close or j > 5:
obj_pose = np.random.uniform(np.array([0, -std_z, -max_y, -std_angle]),
np.array([max_x, std_z, max_y, std_angle])) \
+ np.array([0., avg_z, 0., avg_angle])
if obj_pose_flat:
min_distance = np.linalg.norm(np.array(obj_pose_flat)[:, :3] - obj_pose[:3], axis=1).min()
if min_distance > thresh:
obj_pose_flat.append(obj_pose)
close = False
else:
obj_pose_flat.append(obj_pose)
close = False
obj_id = np.random.choice(np.array(obj_meta)[1:, 0])
indexing = np.argwhere(np.array(obj_meta)[:, 0] == obj_id)[0,0]
obj_classe = obj_meta[indexing, 4]
obj_i = np.concatenate([obj_pose, indexing[None], obj_classe[None]], axis=0)
obj_flat.append(obj_i)
obj_i = tf.cast(np.array(obj_flat), tf.float32)
render_set.append([obj_i])
elif manipulation == 'traffic_jam':
obj_flat = np.reshape(obj, [-1, 6])
obj_ids = np.unique(obj_flat[:, 4])[1:]
np.random.shuffle(obj_flat)
i = 0
while i < len(obj_flat):
obj_i_pose = obj_flat[i, [0, 2]]
# id_i = int(np.random.choice(obj_ids))
# obj_flat[i, 4] = id_i
id_i = int(obj_flat[i, 4])
thresh_close = np.array(id_i).max() * 1.5
close_poses = []
i += 1
for k, pose_k in enumerate(obj_flat[i:]):
x_k, y_k = pose_k[[0, 2]]
d_i_k = np.sqrt((x_k - obj_i_pose[0])**2
+ (y_k - obj_i_pose[1])**2)
if d_i_k < thresh_close:
close_poses.append(i+k)
obj_flat = np.delete(obj_flat, np.array(close_poses), axis=0)
obj_flat = np.delete(obj_flat, tf.where(obj_flat[:, 0] == -1), axis=0)
obj_i = tf.cast(obj_flat, tf.float32)
render_set.append([obj_i])
elif manipulation == 'background':
obj_i = np.reshape(obj[i], [-1, 6])
obj_i = -1 * np.ones_like(obj_i)
obj_i[..., 4] = 0.
obj_i = tf.cast(obj_i, tf.float32)
render_set.append([obj_i])
elif manipulation == 'translate':
k = 15
obj_old = np.reshape(obj[k], [-1, 6])
print(obj_old)
for j in range(len(obj_old)):
for d in np.linspace(-2., 2., 15):
obj_i = -1 * np.ones_like(obj_old)
obj_i[0] = obj_old[j] + np.array([0., 0., d, 0., 0., 0.])
obj_i = tf.cast(obj_i, tf.float32)
render_set.append([obj_i])
return render_set