forked from simo23/tinyYOLOv2
-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_webcam.py
276 lines (194 loc) · 9.43 KB
/
test_webcam.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
import tensorflow as tf
import os
import numpy as np
import net
import weights_loader
import cv2
import warnings
import sys
import time
warnings.filterwarnings('ignore')
def sigmoid(x):
return 1. / (1. + np.exp(-x))
def softmax(x):
e_x = np.exp(x - np.max(x))
out = e_x / e_x.sum()
return out
def iou(boxA,boxB):
# boxA = boxB = [x1,y1,x2,y2]
# Determine the coordinates of the intersection rectangle
xA = max(boxA[0], boxB[0])
yA = max(boxA[1], boxB[1])
xB = min(boxA[2], boxB[2])
yB = min(boxA[3], boxB[3])
# Compute the area of intersection
intersection_area = (xB - xA + 1) * (yB - yA + 1)
# Compute the area of both rectangles
boxA_area = (boxA[2] - boxA[0] + 1) * (boxA[3] - boxA[1] + 1)
boxB_area = (boxB[2] - boxB[0] + 1) * (boxB[3] - boxB[1] + 1)
# Compute the IOU
iou = intersection_area / float(boxA_area + boxB_area - intersection_area)
return iou
def non_maximal_suppression(thresholded_predictions,iou_threshold):
nms_predictions = []
nms_predictions.append(thresholded_predictions[0])
# thresholded_predictions[0] = [x1,y1,x2,y2]
i = 1
while i < len(thresholded_predictions):
n_boxes_to_check = len(nms_predictions)
#print('N boxes to check = {}'.format(n_boxes_to_check))
to_delete = False
j = 0
while j < n_boxes_to_check:
curr_iou = iou(thresholded_predictions[i][0],nms_predictions[j][0])
if(curr_iou > iou_threshold ):
to_delete = True
#print('Checking box {} vs {}: IOU = {} , To delete = {}'.format(thresholded_predictions[i][0],nms_predictions[j][0],curr_iou,to_delete))
j = j+1
if to_delete == False:
nms_predictions.append(thresholded_predictions[i])
i = i+1
return nms_predictions
def preprocessing(input_image,input_height,input_width):
resized_image = cv2.resize(input_image,(input_height, input_width), interpolation = cv2.INTER_CUBIC)
image_data = np.array(resized_image, dtype='f')
# Normalization [0,255] -> [0,1]
image_data /= 255.
# BGR -> RGB? The results do not change much
# copied_image = image_data
#image_data[:,:,2] = copied_image[:,:,0]
#image_data[:,:,0] = copied_image[:,:,2]
image_array = np.expand_dims(image_data, 0) # Add batch dimension
return image_array
def postprocessing(predictions,input_img,score_threshold,iou_threshold,input_height,input_width):
input_image = cv2.resize(input_img,(input_height, input_width), interpolation = cv2.INTER_CUBIC)
n_classes = 20
n_grid_cells = 13
n_b_boxes = 5
n_b_box_coord = 4
# Names and colors for each class
classes = ["aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"]
colors = [(254.0, 254.0, 254), (239.88888888888889, 211.66666666666669, 127),
(225.77777777777777, 169.33333333333334, 0), (211.66666666666669, 127.0, 254),
(197.55555555555557, 84.66666666666667, 127), (183.44444444444443, 42.33333333333332, 0),
(169.33333333333334, 0.0, 254), (155.22222222222223, -42.33333333333335, 127),
(141.11111111111111, -84.66666666666664, 0), (127.0, 254.0, 254),
(112.88888888888889, 211.66666666666669, 127), (98.77777777777777, 169.33333333333334, 0),
(84.66666666666667, 127.0, 254), (70.55555555555556, 84.66666666666667, 127),
(56.44444444444444, 42.33333333333332, 0), (42.33333333333332, 0.0, 254),
(28.222222222222236, -42.33333333333335, 127), (14.111111111111118, -84.66666666666664, 0),
(0.0, 254.0, 254), (-14.111111111111118, 211.66666666666669, 127)]
# Pre-computed YOLOv2 shapes of the k=5 B-Boxes
anchors = [1.08,1.19, 3.42,4.41, 6.63,11.38, 9.42,5.11, 16.62,10.52]
thresholded_predictions = []
#print('Thresholding on (Objectness score)*(Best class score) with threshold = {}'.format(score_threshold))
# IMPORTANT: reshape to have shape = [ 13 x 13 x (5 B-Boxes) x (4 Coords + 1 Obj score + 20 Class scores ) ]
predictions = np.reshape(predictions,(13,13,5,25))
# IMPORTANT: Compute the coordinates and score of the B-Boxes by considering the parametrization of YOLOv2
for row in range(n_grid_cells):
for col in range(n_grid_cells):
for b in range(n_b_boxes):
tx, ty, tw, th, tc = predictions[row, col, b, :5]
# IMPORTANT: (416 img size) / (13 grid cells) = 32!
# YOLOv2 predicts parametrized coordinates that must be converted to full size
# final_coordinates = parametrized_coordinates * 32.0 ( You can see other EQUIVALENT ways to do this...)
center_x = (float(col) + sigmoid(tx)) * 32.0
center_y = (float(row) + sigmoid(ty)) * 32.0
roi_w = np.exp(tw) * anchors[2*b + 0] * 32.0
roi_h = np.exp(th) * anchors[2*b + 1] * 32.0
final_confidence = sigmoid(tc)
# Find best class
class_predictions = predictions[row, col, b, 5:]
class_predictions = softmax(class_predictions)
class_predictions = tuple(class_predictions)
best_class = class_predictions.index(max(class_predictions))
best_class_score = class_predictions[best_class]
# Flip the coordinates on both axes
left = int(center_x - (roi_w/2.))
right = int(center_x + (roi_w/2.))
top = int(center_y - (roi_h/2.))
bottom = int(center_y + (roi_h/2.))
if( (final_confidence * best_class_score) > score_threshold):
thresholded_predictions.append([[left,top,right,bottom],final_confidence * best_class_score,classes[best_class]])
# Sort the B-boxes by their final score
thresholded_predictions.sort(key=lambda tup: tup[1],reverse=True)
#print('Printing {} B-boxes survived after score thresholding:'.format(len(thresholded_predictions)))
#for i in range(len(thresholded_predictions)):
#print('B-Box {} : {}'.format(i+1,thresholded_predictions[i]))
# Non maximal suppression
#print('Non maximal suppression with iou threshold = {}'.format(iou_threshold))
nms_predictions = []
if(len(thresholded_predictions)>0):
nms_predictions = non_maximal_suppression(thresholded_predictions,iou_threshold)
# Print survived b-boxes
#print('Printing the {} B-Boxes survived after non maximal suppression:'.format(len(nms_predictions)))
#for i in range(len(nms_predictions)):
#print('B-Box {} : {}'.format(i+1,nms_predictions[i]))
# Draw final B-Boxes and label on input image
for i in range(len(nms_predictions)):
color = colors[classes.index(nms_predictions[i][2])]
best_class_name = nms_predictions[i][2]
# Put a class rectangle with B-Box coordinates and a class label on the image
input_image = cv2.rectangle(input_image,(nms_predictions[i][0][0],nms_predictions[i][0][1]),(nms_predictions[i][0][2],nms_predictions[i][0][3]),color)
cv2.putText(input_image,best_class_name,(int((nms_predictions[i][0][0]+nms_predictions[i][0][2])/2),int((nms_predictions[i][0][1]+nms_predictions[i][0][3])/2)),cv2.FONT_HERSHEY_SIMPLEX,1,color,3)
return input_image
def inference(sess,preprocessed_image):
# Forward pass of the preprocessed image into the network defined in the net.py file
predictions = sess.run(net.o9,feed_dict={net.x:preprocessed_image})
return predictions
### MAIN ##############################################################################################################
def main(_):
# Definition of the paths
weights_path = './tiny-yolo-voc.weights'
output_image_path = './output.jpg'
ckpt_folder_path = './ckpt/'
# Definition of the parameters
input_height = 416
input_width = 416
score_threshold = 0.3
iou_threshold = 0.3
# Definition of the session
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
# Check for an existing checkpoint and load the weights (if it exists) or do it from binary file
print('Looking for a checkpoint...')
saver = tf.train.Saver()
_ = weights_loader.load(sess,weights_path,ckpt_folder_path,saver)
video_capture = cv2.VideoCapture(0)
n_frames = 0
seconds = 0.0
fps = 0.0
while True:
# Start time
start = time.time()
# Capture frame-by-frame
_, frame = video_capture.read()
n_frames = n_frames + 1
# Preprocess the input image
#print('Preprocessing...')
preprocessed_image = preprocessing(frame,input_height,input_width)
# Compute the predictions on the input image
#print('Computing predictions...')
predictions = []
predictions = inference(sess,preprocessed_image)
# Postprocess the predictions and save the output image
#print('Postprocessing...')
output_image = postprocessing(predictions,frame,score_threshold,iou_threshold,input_height,input_width)
# End time
end = time.time()
# Time elapsed
seconds = (end - start)
# Calculate frames per second
fps = ( fps + (1/seconds) ) / 2
# Display the resulting frame with fps
cv2.putText(output_image,str(fps),(10,50),cv2.FONT_HERSHEY_SIMPLEX,1,(0,0,255),3)
cv2.imshow('Video', output_image)
# Press Q to stop!
if cv2.waitKey(1) & 0xFF == ord('q'):
break
# When everything is done, release the capture
video_capture.release()
cv2.destroyAllWindows()
if __name__ == '__main__':
tf.app.run(main=main)
#######################################################################################################################