-
Notifications
You must be signed in to change notification settings - Fork 3
/
utils.py
234 lines (204 loc) · 8.31 KB
/
utils.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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
import numpy as np
import torch
import cv2
def mask_points_by_range(points, limit_range):
mask = (points[:, 0] >= limit_range[0]) & (points[:, 0] <= limit_range[3]) \
& (points[:, 1] >= limit_range[1]) & (points[:, 1] <= limit_range[4])
return mask
def naive_3diou(keep_box, res_boxes):
'''
keep_box: (7, ), x,y,z,dx,dy,dz,heading
res_boxes: (n, 7)
'''
keep_box_min_x = keep_box[0] - keep_box[3] / 2 # scalar
keep_box_max_x = keep_box[0] + keep_box[3] / 2
res_boxes_min_x = res_boxes[:, 0] - res_boxes[:, 3] / 2 # (n, )
res_boxes_max_x = res_boxes[:, 0] + res_boxes[:, 3] / 2
min_x = np.maximum(res_boxes_min_x, keep_box_min_x) # (n, )
max_x = np.minimum(res_boxes_max_x, keep_box_max_x) # (n, )
x_overlap = max_x - min_x # (n, )
# y
keep_box_min_y = keep_box[1] - keep_box[4] / 2
keep_box_max_y = keep_box[1] + keep_box[4] / 2
res_boxes_min_y = res_boxes[:, 1] - res_boxes[:, 4] / 2
res_boxes_max_y = res_boxes[:, 1] + res_boxes[:, 4] / 2
min_y = np.maximum(res_boxes_min_y, keep_box_min_y)
max_y = np.minimum(res_boxes_max_y, keep_box_max_y)
y_overlap = max_y - min_y
# z
keep_box_min_z = keep_box[2] - keep_box[5] / 2
keep_box_max_z = keep_box[2] + keep_box[5] / 2
res_boxes_min_z = res_boxes[:, 2] - res_boxes[:, 5] / 2
res_boxes_max_z = res_boxes[:, 2] + res_boxes[:, 5] / 2
min_z = np.maximum(res_boxes_min_z, keep_box_min_z)
max_z = np.minimum(res_boxes_max_z, keep_box_max_z)
z_overlap = max_z - min_z
overlap_volumn = x_overlap * y_overlap * z_overlap # (n, )
keep_box_volumn = (keep_box_max_x - keep_box_min_x) * (keep_box_max_y - keep_box_min_y) * (
keep_box_max_z - keep_box_min_z) # scalar
res_boxes_volumn = (res_boxes_max_x - res_boxes_min_x) * (res_boxes_max_y - res_boxes_min_y) * (
res_boxes_max_z - res_boxes_min_z) # (n, )
total_volumn = (keep_box_volumn + res_boxes_volumn) - overlap_volumn # (n, )
ious = overlap_volumn / total_volumn # (n, )
return ious
def check_numpy_to_torch(x):
if isinstance(x, np.ndarray):
return torch.from_numpy(x).float(), True
return x, False
def rotate_points_along_z(points, angle):
"""
Args:
points: (B, N, 3 + C)
angle: (B), angle along z-axis, angle increases x ==> y
Returns:
"""
points, is_numpy = check_numpy_to_torch(points)
angle, _ = check_numpy_to_torch(angle)
cosa = torch.cos(angle)
sina = torch.sin(angle)
zeros = angle.new_zeros(points.shape[0])
ones = angle.new_ones(points.shape[0])
rot_matrix = torch.stack((
cosa, sina, zeros,
-sina, cosa, zeros,
zeros, zeros, ones
), dim=1).view(-1, 3, 3).float()
points_rot = torch.matmul(points[:, :, 0:3], rot_matrix)
points_rot = torch.cat((points_rot, points[:, :, 3:]), dim=-1)
return points_rot.numpy() if is_numpy else points_rot
def boxes_to_corners_3d(boxes3d):
"""
7 -------- 4
/| /|
6 -------- 5 .
| | | |
. 3 -------- 0
|/ |/
2 -------- 1
Args:
boxes3d: (N, 7) [x, y, z, dx, dy, dz, heading], (x, y, z) is the box center
Returns:
"""
boxes3d, is_numpy = check_numpy_to_torch(boxes3d)
template = boxes3d.new_tensor((
[1, 1, -1], [1, -1, -1], [-1, -1, -1], [-1, 1, -1],
[1, 1, 1], [1, -1, 1], [-1, -1, 1], [-1, 1, 1],
)) / 2
corners3d = boxes3d[:, None, 3:6].repeat(1, 8, 1) * template[None, :, :]
corners3d = rotate_points_along_z(corners3d.view(-1, 8, 3), boxes3d[:, 6]).view(-1, 8, 3)
corners3d += boxes3d[:, None, 0:3]
return corners3d.numpy() if is_numpy else corners3d
def draw_mask(corners, blank_map):
'''
corners: (8, 3), x,y,z
blank_map: (H, W)
'''
# contours: (n,1,2)
contours = corners[:4, :2] # (4, 2), float
contours = contours.astype(np.int32)
H, W = blank_map.shape
contours[:, 0] = np.clip(contours[:, 0], 0, W)
contours[:, 1] = np.clip(contours[:, 1], 0, H)
contours = np.expand_dims(contours, axis=1) # (4, 1, 2), float
img = blank_map.copy()
cv2.drawContours(img, [contours], -1, 1, -1)
return img.astype(np.bool)
def mask_3diou(keep_box, res_boxes, scale_factor=100):
'''
keep_box: (7, ), x,y,z,dx,dy,dz,heading
res_boxes: (n, 7), x,y,z,dx,dy,dz,heading
scale_factor: enlarge points coordinates to transform them into voxels
returns:
ious: (n, )
'''
keep_box = keep_box.copy()
res_boxes = res_boxes.copy()
keep_box[: 6] *= scale_factor
res_boxes[:, :6] *= scale_factor
all_boxes = np.concatenate((np.expand_dims(keep_box, axis=0), res_boxes), axis=0) # (n+1, 7)
all_corners = boxes_to_corners_3d(all_boxes) # (n+1, 8, 3)
all_min_x = np.min(all_corners[:, :, 0])
all_max_x = np.max(all_corners[:, :, 0])
all_min_y = np.min(all_corners[:, :, 1])
all_max_y = np.max(all_corners[:, :, 1])
H = int(all_max_y - all_min_y)
W = int(all_max_x - all_min_x)
all_corners[:, :, :2] -= [all_min_x, all_min_y]
keep_corners = all_corners[0, :, :] # (8, 3)
res_corners = all_corners[1:, :, :] # (n, 8, 3)
blank_map = np.zeros((H, W))
keep_mask = draw_mask(keep_corners, blank_map)
# plt.imshow(keep_mask)
# plt.show()
# z
keep_box_min_z = keep_box[2] - keep_box[5] / 2
keep_box_max_z = keep_box[2] + keep_box[5] / 2
res_boxes_min_z = res_boxes[:, 2] - res_boxes[:, 5] / 2
res_boxes_max_z = res_boxes[:, 2] + res_boxes[:, 5] / 2
min_z = np.maximum(res_boxes_min_z, keep_box_min_z) # overlap area
max_z = np.minimum(res_boxes_max_z, keep_box_max_z)
z_overlap = max_z - min_z # (n, )
ious = np.zeros(len(z_overlap)) # (n, )
for i in range(len(z_overlap)):
h_i = z_overlap[i]
mask_i = draw_mask(res_corners[i, :, :], blank_map)
overlap_i = h_i * (np.sum(mask_i & keep_mask))
union = np.sum(keep_mask) * keep_box[5] + np.sum(mask_i) * res_boxes[i, 5]
ious[i] = overlap_i / union
return ious
def nms_3d(boxes, scores, nms_thres=0.1, score_thres=0.6):
'''
boxes: (n, 7), x,y,z,dx,dy,dz,heading
scores: (n, )
returns:
filtered_boxes: (n2, 7)
keep_inds: indices of filtered boxes
'''
sorted_inds = np.argsort(-scores)
keep_inds = []
while len(sorted_inds)>0:
if scores[sorted_inds[0]] < score_thres:
break
keep_inds.append(sorted_inds[0])
if len(sorted_inds) == 1:
break
keep_box = boxes[sorted_inds[0], :]
res_inds = sorted_inds[1:]
res_boxes = boxes[res_inds, :]
se_inds = np.arange(len(res_inds))
# x
keep_box_min_x = keep_box[0] - keep_box[3]/2 # scalar
keep_box_max_x = keep_box[0] + keep_box[3]/2
res_boxes_min_x = res_boxes[:, 0] - res_boxes[:, 3]/2 # (n, )
res_boxes_max_x = res_boxes[:, 0] + res_boxes[:, 3]/2
min_x = np.maximum(res_boxes_min_x, keep_box_min_x) # (n, )
max_x = np.minimum(res_boxes_max_x, keep_box_max_x) # (n, )
x_overlap = max_x > min_x
# y
keep_box_min_y = keep_box[1] - keep_box[4]/2
keep_box_max_y = keep_box[1] + keep_box[4]/2
res_boxes_min_y = res_boxes[:, 1] - res_boxes[:, 4]/2
res_boxes_max_y = res_boxes[:, 1] + res_boxes[:, 4]/2
min_y = np.maximum(res_boxes_min_y, keep_box_min_y)
max_y = np.minimum(res_boxes_max_y, keep_box_max_y)
y_overlap = max_y > min_y
# z
keep_box_min_z = keep_box[2] - keep_box[5]/2
keep_box_max_z = keep_box[2] + keep_box[5]/2
res_boxes_min_z = res_boxes[:, 2] - res_boxes[:, 5]/2
res_boxes_max_z = res_boxes[:, 2] + res_boxes[:, 5]/2
min_z = np.maximum(res_boxes_min_z, keep_box_min_z)
max_z = np.minimum(res_boxes_max_z, keep_box_max_z)
z_overlap = max_z > min_z
overlap_mask = x_overlap & y_overlap & z_overlap
care_res_boxes = res_boxes[overlap_mask, :] # (m, 7)
overlap_se_inds = se_inds[overlap_mask]
ious = mask_3diou(keep_box, care_res_boxes) # (m, )
delete_mask = ious > nms_thres
delete_se_inds = overlap_se_inds[delete_mask]
sorted_inds = np.delete(res_inds, delete_se_inds)
if len(keep_inds)>0:
keep_inds = np.array(keep_inds)
return scores[keep_inds], keep_inds
else:
return None, None