-
Notifications
You must be signed in to change notification settings - Fork 3
/
evaluate.py
314 lines (259 loc) · 9.55 KB
/
evaluate.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
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
# -*- coding: utf-8 -*-
"""
.. codeauthor:: Daniel Seichter <[email protected]>
"""
import argparse as ap
from functools import partial
import json
import multiprocessing
import os
import warnings
import cv2
import numpy as np
from tqdm import tqdm
from tqdm.contrib.concurrent import thread_map
from nicr_scene_analysis_datasets import Hypersim
from utils import DEFAULT_DATASET_PATH
from utils import DEFAULT_PREDICTIONS_PATH
def confusion_matrix_fast(pred, gt, n_classes):
# note: this function is 15x faster than sklearn.metrics.confusion_matrix
# determine dtype for unique mapping
n_classes_squared = n_classes**2
if n_classes_squared < 2**(8-1)-1:
dtype = np.int8
elif n_classes_squared < 2**(16-1)-1:
dtype = np.int16
else:
dtype = np.int64 # equal to long
# convert to dtype
pred_ = pred.astype(dtype)
gt_ = gt.astype(dtype)
# compute confusion matrix
unique_mapping = (gt_.reshape(-1)*n_classes + pred_.reshape(-1))
cnts = np.bincount(unique_mapping,
minlength=n_classes_squared)
return cnts.reshape(n_classes, n_classes)
def get_confusion_matrix_for_sample(
sample_idx,
dataset,
prediction_basepath,
prediction_extension='.png',
prediction_contains_void=True,
max_depth_in_m=20 # max 20m
):
n_classes = dataset.semantic_n_classes # with void
# get sample
sample = dataset[sample_idx]
# load prediction
fp = os.path.join(prediction_basepath, *sample['identifier'])
fp += prediction_extension
if '.png' == prediction_extension:
# prediction is given as image
pred = cv2.imread(fp, cv2.IMREAD_UNCHANGED)
if pred is None:
raise IOError(f"Cannot load '{fp}'")
if pred.ndim > 2:
warnings.warn(f"Prediction ('{fp}') has more than one channel. "
"Using first channel.")
pred = pred[..., 0]
elif '.npy' == prediction_extension:
# prediction is given as numpy array with shape (h, w, topk)
pred = np.load(fp)
pred = pred[0, ...].astype('uint8') # use top1 only
if not prediction_contains_void:
pred += 1
# create flat views
gt = sample['semantic'].reshape(-1)
pred = pred.reshape(-1)
# mask using max depth
if max_depth_in_m is not None:
depth = sample['depth'].reshape(-1)
mask = depth < (max_depth_in_m*1000)
gt = gt[mask]
pred = pred[mask]
# move invalid pixels in prediction, i.e., pixels that may indicate free
# space, to class with index i=n_classes
pred[pred > (n_classes-1)] = n_classes
n_classes = n_classes + 1 # +1 = invalid pixels
return confusion_matrix_fast(pred, gt, n_classes=n_classes)
def get_measures(cm, ignore_void=True):
# cm is gt x pred with void + n_classes + invalid (free space)
tp = np.diag(cm)
sum_gt = cm.sum(axis=1)
sum_pred = cm.sum(axis=0)
invalid_pixels = cm[:, -1]
if ignore_void:
# void is first class (idx=0)
tp = tp[1:]
sum_pred = sum_pred[1:]
sum_gt = sum_gt[1:]
sum_pred -= cm[0, 1:] # do not count fp for void
invalid_pixels = invalid_pixels[1:]
n_total_pixels = sum_gt.sum()
# we do want ignore classes without gt pixels
gt_mask = sum_gt != 0
# invalid pixels
invalid_ratio = invalid_pixels.sum() / n_total_pixels
with np.errstate(divide='ignore', invalid='ignore'):
invalid_ratios = invalid_pixels / sum_gt
invalid_mean_ratio_gt_masked = np.mean(invalid_ratios[gt_mask])
valid_weights = 1 - invalid_ratios
# intersection over union
intersections = tp
unions = sum_pred + sum_gt - tp
with np.errstate(divide='ignore', invalid='ignore'):
ious = intersections / unions.astype(np.float32)
# mean intersection over union and gt masked version
miou = np.mean(np.nan_to_num(ious, nan=0.0))
miou_gt_masked = np.mean(ious[gt_mask])
# frequency weighted intersection over union
# normal fwiou and gt masked version are equal
fwiou_gt_masked = np.sum(ious[gt_mask] * tp[gt_mask]/n_total_pixels)
# pixel accuracy and mean pixel accuracy
pacc = tp.sum() / sum_gt.sum()
with np.errstate(divide='ignore', invalid='ignore'):
paccs = tp / sum_gt
mean_pacc_gt_masked = np.mean(tp[gt_mask] / sum_gt[gt_mask])
# valid weighted mean intersection over union
vwmiou_gt_masked = np.mean(ious[gt_mask]*valid_weights[gt_mask])
# valid weighted mean pixel accuracy
vwmean_pacc_gt_masked = np.mean(tp[gt_mask] / sum_gt[gt_mask] * valid_weights[gt_mask])
# build dict of measures
measures = {
'cm': cm.tolist(),
'invalid_ratio': invalid_ratio,
'invalid_ratios': invalid_ratios.tolist(),
'invalid_mean_ratio_gt_masked': invalid_mean_ratio_gt_masked,
'ious': ious.tolist(),
'miou': miou,
'miou_gt_masked': miou_gt_masked,
'fwiou_gt_masked': fwiou_gt_masked,
'pacc': pacc,
'paccs': paccs.tolist(),
'mean_pacc_gt_masked': mean_pacc_gt_masked,
'vwmiou_gt_masked': vwmiou_gt_masked,
'vwmean_pacc_gt_masked': vwmean_pacc_gt_masked,
}
return measures
def _parse_args():
parser = ap.ArgumentParser(formatter_class=ap.ArgumentDefaultsHelpFormatter)
parser.add_argument(
'--dataset-path',
type=str,
default=DEFAULT_DATASET_PATH,
help="Path to the dataset."
)
parser.add_argument(
'--dataset-split',
type=str,
default='test',
help="Dataset split to use."
)
parser.add_argument(
'--predictions-path',
type=str,
default=DEFAULT_PREDICTIONS_PATH,
help="Path to stored predicted semantic segmentation. Use an empty "
"string to skip the evaluating the predicted semantic "
"segmentation."
)
parser.add_argument(
'--result-paths',
nargs='+',
type=str,
help="Paths to further results.",
default=[]
)
parser.add_argument(
'--force-recomputing',
action='store_true',
default=False,
help="Force recomputing."
)
parser.add_argument(
'--n-worker',
type=int,
default=min(multiprocessing.cpu_count(), 48),
help="Number of workers to use."
)
return parser.parse_args()
def main():
# args
args = _parse_args()
# just obtain all sample names
dataset = Hypersim(dataset_path=args.dataset_path,
split=args.dataset_split,
subsample=None,
sample_keys=('identifier',))
samples = [s['identifier'] for s in dataset] # tuple (scene, cam, id)
scenes = sorted(list(set(s[0] for s in samples)))
# load dataset
dataset = Hypersim(dataset_path=args.dataset_path,
split=args.dataset_split,
subsample=None,
sample_keys=('identifier', 'depth', 'semantic'),
use_cache=False,
cache_disable_deepcopy=False)
# get paths to evaluate
paths = []
if args.predictions_path:
# evaluate the network prediction
paths += [
os.path.join(args.predictions_path, args.dataset_split,
Hypersim.SEMANTIC_DIR),
]
paths += [
os.path.join(path) for path in args.result_paths
]
# run evaluation
for path in tqdm(paths):
print(f"Evaluating: '{path}'")
results_fp = os.path.join(path, 'results.json')
if os.path.exists(results_fp) and not args.force_recomputing:
continue
# get confusion matrices
if 1 == args.n_worker:
cms = []
for i in tqdm(range(len(dataset))):
cm = get_confusion_matrix_for_sample(
i,
dataset=dataset,
prediction_basepath=path,
prediction_extension='.png',
prediction_contains_void=True,
max_depth_in_m=20
)
cms.append(cm)
else:
f = partial(get_confusion_matrix_for_sample,
dataset=dataset,
prediction_basepath=path,
prediction_extension='.png',
prediction_contains_void=True,
max_depth_in_m=20)
cms = thread_map(f, list(range(len(dataset))),
max_workers=args.n_worker,
chunksize=10,
leave=False)
# get overall measures
assert len(cms) == len(samples)
cm = np.array(cms).sum(axis=0)
measures = get_measures(cm, ignore_void=True)
for k in ('miou_gt_masked', 'mean_pacc_gt_masked',
'invalid_ratio', 'invalid_mean_ratio_gt_masked',
'vwmiou_gt_masked', 'vwmean_pacc_gt_masked'):
print(f"{k}: {measures[k]}")
# get results for each scene
cms_per_scene = {s: [] for s in scenes}
for cm, sample in zip(cms, samples):
scene = sample[0]
cms_per_scene[scene].append(cm)
measures['per_scene'] = {}
for scene, cms_scene in cms_per_scene.items():
cm = np.array(cms_scene).sum(axis=0)
measures['per_scene'][scene] = get_measures(cm, ignore_void=True)
# write results to file
with open(results_fp, 'w') as f:
json.dump(measures, f, indent=4)
if __name__ == '__main__':
main()