-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtest_perception_performance.py
182 lines (150 loc) · 6.15 KB
/
test_perception_performance.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
import h5py, cv2, sys, pickle, math, numpy
import numpy as np
sys.path.append("/data/yang/code/aws")
from all_perceptions import Perceptions
class SegPerformance():
def __init__(self):
self.p = Perceptions(det_COCO=False,
det_TL=False,
det_TS=False,
seg=True,
depth=False,
batch_size=1, # batch_size could also be a dict
gpu_assignment=[0],
compute_methods={},
viz_methods={},
num_replicates={},
path_config="path_jormungandr")
self.accu = np.zeros((6, 2), dtype=np.int64)
def _compute_inter_union(self, gt, pred, nclasses):
# return a nclass * 2 matrix: [iclass][inter, union]
out = np.zeros((nclasses, 2), dtype=np.int64)
pred = pred[gt != 0] - 1
gt = gt[gt != 0] - 1
for i in range(nclasses):
out[i, 0] = np.sum(np.logical_and(pred == i, gt == i))
out[i, 1] = np.sum(np.logical_or(pred == i, gt == i))
return out
def reset(self):
self.accu = np.zeros((6, 2), dtype=np.int64)
def one_pair(self, image, seg):
'''
0 None
1 Buildings
2 Fences
3 Other
4 Pedestrians
5 Poles
6 RoadLines
7 Roads
8 Sidewalks
9 Vegetation
10 Vehicles
11 Walls
12 TrafficSigns
# 0 1 2 3 4 5 6
{'Ignored', 'Movable', 'Navigable', 'NoneNavigable', 'StaticLayout', 'Sky', 'Lane'}
'''
image_batch = np.expand_dims(image, 0)
out = self.p.compute(image_batch)
out = np.squeeze(out['seg'])
out = np.argmax(out, axis=2) + 1 # 0 does not exists, since it is the ignore class underhood
out = cv2.resize(out, (out.shape[1] * 4, out.shape[0] * 4), interpolation=cv2.INTER_NEAREST)
mapping = np.array([0, 4, 3, 0, 1, 4, 6, 2, 3, 4, 1, 4, 4])
seg_cvt = mapping[seg]
self.accu += self._compute_inter_union(seg_cvt, out, 6)
def summarize(self):
names = ['Movable', 'Navigable', 'NoneNavigable', 'StaticLayout', 'Sky', 'Lane']
out = {}
for i in range(len(names)):
out[names[i]] = 1.0 * self.accu[i, 0] / self.accu[i, 1]
return out
class DepthPerformance():
def __init__(self):
width = 768
fov = 90.0
self.focal = width / (2.0 * math.tan(fov * math.pi / 360.0))
self.baseline = 0.54
self.p = Perceptions(det_COCO=False,
det_TL=False,
det_TS=False,
seg=False,
depth=True,
batch_size=1, # batch_size could also be a dict
gpu_assignment=[0],
compute_methods={},
viz_methods={},
num_replicates={},
path_config="path_jormungandr")
self.reset()
def _compute_errors(self, gt, pred):
thresh = np.maximum((gt / pred), (pred / gt))
a1 = (thresh < 1.25).mean()
a2 = (thresh < 1.25 ** 2).mean()
a3 = (thresh < 1.25 ** 3).mean()
rmse = (gt - pred) ** 2
rmse = np.sqrt(rmse.mean())
rmse_log = (np.log(gt) - np.log(pred)) ** 2
rmse_log = np.sqrt(rmse_log.mean())
abs_rel = np.mean(np.abs(gt - pred) / gt)
sq_rel = np.mean(((gt - pred) ** 2) / gt)
return abs_rel, sq_rel, rmse, rmse_log, a1, a2, a3
def reset(self):
self.accu = []
self.count = 0
def one_pair(self, image, depth):
image_batch = np.expand_dims(image, 0)
out = self.p.compute(image_batch)
out = np.squeeze(out['depth'])
out = cv2.resize(out, (image.shape[1], image.shape[0]), interpolation=cv2.INTER_LINEAR)
out = self.baseline * self.focal / (out + 0.0000001)
out = out / 2000.0 # some magic number that I made up
errors = self._compute_errors(depth + 1e-5, out + 1e-5)
self.accu.append(errors)
self.count += 1
def summarize(self):
names = ['abs_rel', 'sq_rel', 'rmse', 'rmse_log', 'a1', 'a2', 'a3']
errors = np.stack(self.accu, axis=0)
errors = np.mean(errors, 0)
out = {}
for i in range(len(names)):
out[names[i]] = errors[i]
return out
modality = "depth" # "seg"
perf = {"depth": DepthPerformance, "seg": None}[modality]()
results = {}
for town in ["Town01", "Town02"]:
results[town] = {}
for weather in range(1, 14):
perf.reset()
path = "/scratch/yang/aws_data/carla_collect/" + town + "_allsensor/default_RotationPitch=0_WeatherId=" + str(
weather).zfill(2) + "/data_00000.h5"
f = h5py.File(path, "r")
for i in range(200):
# print(i)
image = f['CameraMiddle'][i]
image = cv2.imdecode(image, 1)
image = image[:, :, ::-1]
if modality == "seg":
seg = f['SegMiddle'][i]
mode_image = cv2.imdecode(seg, 0)
elif modality == "depth":
mode_image = f['DepthMiddle'][i]
mode_image = cv2.imdecode(mode_image, 1)
def to_depth(array):
# input is bgr
array = array.astype(numpy.float32)
# Apply (R + G * 256 + B * 256 * 256) / (256 * 256 * 256 - 1).
normalized_depth = numpy.dot(array[:, :, :], [65536.0, 256.0, 1.0])
normalized_depth /= 16777215.0 / 1000.0 # (256.0 * 256.0 * 256.0 - 1.0)
return normalized_depth
mode_image = to_depth(mode_image)
else:
raise
perf.one_pair(image, mode_image)
f.close()
print(town, "weather", weather, perf.summarize())
results[town][weather] = perf.summarize()
with open(modality + ".pkl", "w") as ppp:
pickle.dump(results, ppp)
print("done")