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
from abc import ABCMeta, abstractmethod
import numpy as np
import sys
import argparse
import time
from imagenet_cls_test_alexnet import CaffeModel, DnnCaffeModel
try:
import cv2 as cv
except ImportError:
raise ImportError('Can\'t find OpenCV Python module. If you\'ve built it from sources without installation, '
'configure environment variable PYTHONPATH to "opencv_build_dir/lib" directory (with "python3" subdirectory if required)')
def get_metrics(conf_mat):
pix_accuracy = np.trace(conf_mat) / np.sum(conf_mat)
t = np.sum(conf_mat, 1)
num_cl = np.count_nonzero(t)
assert num_cl
mean_accuracy = np.sum(np.nan_to_num(np.divide(np.diagonal(conf_mat), t))) / num_cl
col_sum = np.sum(conf_mat, 0)
mean_iou = np.sum(
np.nan_to_num(np.divide(np.diagonal(conf_mat), (t + col_sum - np.diagonal(conf_mat))))) / num_cl
return pix_accuracy, mean_accuracy, mean_iou
def eval_segm_result(net_out):
assert type(net_out) is np.ndarray
assert len(net_out.shape) == 4
channels_dim = 1
y_dim = channels_dim + 1
x_dim = y_dim + 1
res = np.zeros(net_out.shape).astype(np.int)
for i in range(net_out.shape[y_dim]):
for j in range(net_out.shape[x_dim]):
max_ch = np.argmax(net_out[..., i, j])
res[0, max_ch, i, j] = 1
return res
def get_conf_mat(gt, prob):
assert type(gt) is np.ndarray
assert type(prob) is np.ndarray
conf_mat = np.zeros((gt.shape[0], gt.shape[0]))
for ch_gt in range(conf_mat.shape[0]):
gt_channel = gt[ch_gt, ...]
for ch_pr in range(conf_mat.shape[1]):
prob_channel = prob[ch_pr, ...]
conf_mat[ch_gt][ch_pr] = np.count_nonzero(np.multiply(gt_channel, prob_channel))
return conf_mat
class MeanChannelsPreproc:
def __init__(self):
pass
@staticmethod
def process(img):
image_data = np.array(img).transpose(2, 0, 1).astype(np.float32)
mean = np.ones(image_data.shape)
mean[0] *= 104
mean[1] *= 117
mean[2] *= 123
image_data -= mean
image_data = np.expand_dims(image_data, 0)
return image_data
class DatasetImageFetch(object):
__metaclass__ = ABCMeta
data_prepoc = object
@abstractmethod
def __iter__(self):
pass
@abstractmethod
def next(self):
pass
@staticmethod
def pix_to_c(pix):
return pix[0] * 256 * 256 + pix[1] * 256 + pix[2]
@staticmethod
def color_to_gt(color_img, colors):
num_classes = len(colors)
gt = np.zeros((num_classes, color_img.shape[0], color_img.shape[1])).astype(np.int)
for img_y in range(color_img.shape[0]):
for img_x in range(color_img.shape[1]):
c = DatasetImageFetch.pix_to_c(color_img[img_y][img_x])
if c in colors:
cls = colors.index(c)
gt[cls][img_y][img_x] = 1
return gt
class PASCALDataFetch(DatasetImageFetch):
img_dir = ''
segm_dir = ''
names = []
colors = []
i = 0
def __init__(self, img_dir, segm_dir, names_file, segm_cls_colors_file, preproc):
self.img_dir = img_dir
self.segm_dir = segm_dir
self.colors = self.read_colors(segm_cls_colors_file)
self.data_prepoc = preproc
self.i = 0
with open(names_file) as f:
for l in f.readlines():
self.names.append(l.rstrip())
@staticmethod
def read_colors(img_classes_file):
result = []
with open(img_classes_file) as f:
for l in f.readlines():
color = np.array(map(int, l.split()[1:]))
result.append(DatasetImageFetch.pix_to_c(color))
return result
def __iter__(self):
return self
def next(self):
if self.i < len(self.names):
name = self.names[self.i]
self.i += 1
segm_file = self.segm_dir + name + ".png"
img_file = self.img_dir + name + ".jpg"
gt = self.color_to_gt(cv.imread(segm_file, cv.IMREAD_COLOR)[:, :, ::-1], self.colors)
img = self.data_prepoc.process(cv.imread(img_file, cv.IMREAD_COLOR)[:, :, ::-1])
return img, gt
else:
self.i = 0
raise StopIteration
def get_num_classes(self):
return len(self.colors)
class SemSegmEvaluation:
log = file
def __init__(self, log_path,):
self.log = open(log_path, 'w')
def process(self, frameworks, data_fetcher):
samples_handled = 0
conf_mats = [np.zeros((data_fetcher.get_num_classes(), data_fetcher.get_num_classes())) for i in range(len(frameworks))]
blobs_l1_diff = [0] * len(frameworks)
blobs_l1_diff_count = [0] * len(frameworks)
blobs_l_inf_diff = [sys.float_info.min] * len(frameworks)
inference_time = [0.0] * len(frameworks)
for in_blob, gt in data_fetcher:
frameworks_out = []
samples_handled += 1
for i in range(len(frameworks)):
start = time.time()
out = frameworks[i].get_output(in_blob)
end = time.time()
segm = eval_segm_result(out)
conf_mats[i] += get_conf_mat(gt, segm[0])
frameworks_out.append(out)
inference_time[i] += end - start
pix_acc, mean_acc, miou = get_metrics(conf_mats[i])
name = frameworks[i].get_name()
print >> self.log, samples_handled, 'Pixel accuracy, %s:' % name, 100 * pix_acc
print >> self.log, samples_handled, 'Mean accuracy, %s:' % name, 100 * mean_acc
print >> self.log, samples_handled, 'Mean IOU, %s:' % name, 100 * miou
print >> self.log, "Inference time, ms ", \
frameworks[i].get_name(), inference_time[i] / samples_handled * 1000
for i in range(1, len(frameworks)):
log_str = frameworks[0].get_name() + " vs " + frameworks[i].get_name() + ':'
diff = np.abs(frameworks_out[0] - frameworks_out[i])
l1_diff = np.sum(diff) / diff.size
print >> self.log, samples_handled, "L1 difference", log_str, l1_diff
blobs_l1_diff[i] += l1_diff
blobs_l1_diff_count[i] += 1
if np.max(diff) > blobs_l_inf_diff[i]:
blobs_l_inf_diff[i] = np.max(diff)
print >> self.log, samples_handled, "L_INF difference", log_str, blobs_l_inf_diff[i]
self.log.flush()
for i in range(1, len(blobs_l1_diff)):
log_str = frameworks[0].get_name() + " vs " + frameworks[i].get_name() + ':'
print >> self.log, 'Final l1 diff', log_str, blobs_l1_diff[i] / blobs_l1_diff_count[i]
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--imgs_dir", help="path to PASCAL VOC 2012 images dir, data/VOC2012/JPEGImages")
parser.add_argument("--segm_dir", help="path to PASCAL VOC 2012 segmentation dir, data/VOC2012/SegmentationClass/")
parser.add_argument("--val_names", help="path to file with validation set image names, download it here: "
"https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/data/pascal/seg11valid.txt")
parser.add_argument("--cls_file", help="path to file with colors for classes, download it here: "
"https://github.com/opencv/opencv/blob/master/samples/data/dnn/pascal-classes.txt")
parser.add_argument("--prototxt", help="path to caffe prototxt, download it here: "
"https://github.com/opencv/opencv/blob/master/samples/data/dnn/fcn8s-heavy-pascal.prototxt")
parser.add_argument("--caffemodel", help="path to caffemodel file, download it here: "
"http://dl.caffe.berkeleyvision.org/fcn8s-heavy-pascal.caffemodel")
parser.add_argument("--log", help="path to logging file")
parser.add_argument("--in_blob", help="name for input blob", default='data')
parser.add_argument("--out_blob", help="name for output blob", default='score')
args = parser.parse_args()
prep = MeanChannelsPreproc()
df = PASCALDataFetch(args.imgs_dir, args.segm_dir, args.val_names, args.cls_file, prep)
fw = [CaffeModel(args.prototxt, args.caffemodel, args.in_blob, args.out_blob, True),
DnnCaffeModel(args.prototxt, args.caffemodel, '', args.out_blob)]
segm_eval = SemSegmEvaluation(args.log)
segm_eval.process(fw, df)