def test_model(self, opt, val_data, weight_path, out_path):
self.model_all.load_weights(weight_path, by_name=True)
print 'load weights from {}'.format(weight_path)
res_all = []
#用来保存预测结果
res_file = os.path.join(out_path, 'val_det.txt')
start_time = time.time()
for f in range(len(val_data)):
filepath = val_data[f]['filepath']
# print (filepath)
frame_number = f + 1
img = cv2.imread(filepath)
x_in = bbox_process.format_img(img, opt)
#Y表示所有stage的分类以及回归的预测结果,2step返回4个值
Y = self.model_all.predict(x_in)
#Y[0],Y[1]表示 alf_1st的预测结果，Y[2],Y[3]表示alf+2nd的预测结果
#Y[0]以及Y[2]表示cls_pred,Y[1]以及Y[3]表示reg_pred
proposals = bbox_process.pred_pp_1st(self.anchors, Y[0], Y[1], opt)
#bbx:[x1,y1,x2,y2]
bbx, scores = bbox_process.pred_det(proposals,
Y[2],
Y[3],
opt,
step=2)
f_res = np.repeat(frame_number, len(bbx), axis=0).reshape((-1, 1))
#generate width and height
bbx[:, [2, 3]] -= bbx[:, [0, 1]]
res_all += np.concatenate((f_res, bbx, scores), axis=-1).tolist()
np.savetxt(res_file, np.array(res_all), fmt='%.4f')
print 'Test time: %.4f s' % (time.time() - start_time)
def test_model(self, opt, val_data, weight_path, out_path):
self.model_all.load_weights(weight_path, by_name=True)
print 'load weights from {}'.format(weight_path)
res_all = []
res_file = os.path.join(out_path, 'val_det.txt')
start_time = time.time()
for f in range(len(val_data)):
filepath = val_data[f]['filepath']
frame_number = f + 1
img = cv2.imread(filepath)
x_in = bbox_process.format_img(img, opt)
Y = self.model_all.predict(x_in)
proposals = bbox_process.pred_pp_1st(self.anchors, Y[0], Y[1], opt)
proposals_2nd = bbox_process.pred_pp_2nd(proposals, Y[2], Y[3],
opt)
bbx, scores = bbox_process.pred_det(proposals_2nd,
Y[4],
Y[5],
opt,
step=3)
f_res = np.repeat(frame_number, len(bbx), axis=0).reshape((-1, 1))
bbx[:, [2, 3]] -= bbx[:, [0, 1]]
res_all += np.concatenate((f_res, bbx, scores), axis=-1).tolist()
np.savetxt(res_file, np.array(res_all), fmt='%.4f')
print 'Test time: %.4f s' % (time.time() - start_time)
def demo_onepic(self, opt, img_path, weight_path, output_path):
self.model_all.load_weights(weight_path, by_name=True)
print 'load weights from {}'.format(weight_path)
img = cv2.imread(img_path)
x_in = bbox_process.format_img(img, opt)
Y = self.model_all.predict(x_in)
# print Y[2]
proposals = bbox_process.pred_pp_1st(self.anchors, Y[0], Y[1], opt)
bbx, scores = bbox_process.pred_det(proposals, Y[2], Y[3], opt, step=2)
for ind in range(len(bbx)):
(x1, y1, x2, y2) = bbx[ind, :]
cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 2)
cv2.imshow('det_result', img)
cv2.waitKey(0)
def demo(self,opt, val_data, weight_path, out_path):
self.model_all.load_weights(weight_path, by_name=True)
print 'load weights from {}'.format(weight_path)
for f in range(len(val_data)):
img_name = os.path.join('data/examples/',val_data[f])
if not img_name.lower().endswith(('.jpg', '.png')):
continue
print(img_name)
img = cv2.imread(img_name)
x_in = bbox_process.format_img(img, opt)
Y = self.model_all.predict(x_in)
proposals = bbox_process.pred_pp_1st(self.anchors, Y[0], Y[1], opt)
bbx, scores = bbox_process.pred_det(proposals, Y[2], Y[3], opt, step=2)
for ind in range(len(bbx)):
(x1, y1, x2, y2) = bbx[ind, :]
cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 2)
cv2.imwrite(os.path.join(out_path, val_data[f]),img)
def test_model_with_input_image(self, img, opt, weight_path):
self.model_all.load_weights(weight_path, by_name=True)
print('load weights from {}'.format(weight_path))
x_in = bbox_process.format_img(img, opt)
print('finished format_img.')
Y = self.model_all.predict(x_in)
print('finished model predict.')
print('begin to gen proposals...')
proposals = bbox_process.pred_pp_1st(self.anchors, Y[0], Y[1], opt)
print('finished proposals generation.')
print('begin to det box & score.')
bbx, scores = bbox_process.pred_det(proposals, Y[2], Y[3], opt, step=2)
# bbx[:, [2, 3]] -= bbx[:, [0, 1]]
print('finished box & score det.')
return bbx, scores
def demo_video(self, opt, video_path, weight_path, output_path):
self.model_all.load_weights(weight_path, by_name=True)
print 'load weights from {}'.format(weight_path)
print video_path
timeF = 3
cap = cv2.VideoCapture(video_path)
ret, frame = cap.read()
#resize成为模型能读取的大小
frame = cv2.resize(frame, (2048, 1024))
#格式转化 ，减去mean
x_in = bbox_process.format_img(frame, opt)
if not cap.isOpened():
raise IOError("Couldn't open webcam or video")
# print (codec,size)
codec = cv2.VideoWriter_fourcc(*'MJPG')
size = ((2048), (1024))
out = cv2.VideoWriter(
output_path,
codec,
15.0, #15 frame per second
size)
c = 1
dets = []
while cap.isOpened():
det = []
ret, frame = cap.read()
if frame is not None:
if (c % timeF == 0):
frame = cv2.resize(frame, (2048, 1024))
x_in = bbox_process.format_img(frame, opt)
Y = self.model_all.predict(x_in)
proposals = bbox_process.pred_pp_1st(
self.anchors, Y[0], Y[1], opt)
bbx, scores = bbox_process.pred_det(proposals,
Y[2],
Y[3],
opt,
step=2)
# dets=np.hstack((bbx,scores[:,np.newaxis])).astype(float64)
for ind in range(len(bbx)):
(x1, y1, x2, y2) = bbx[ind, :]
# print (x1,y1,x2,y2)
# if (x2-x1<100)&(y2-y1<350)&(x2-x1>25)&(y2-y1>150):
# cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
# det.append([x1,y1,x2,y2])
# if (x2-x1<100)&(y2-y1<350)&(x2-x1>25)&(y2-y1>150):
cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0),
2)
# det.append([x1,y1,x2,y2])
# print type(frame)
dets.append(det)
cv2.imshow('detection', frame)
out.write(frame)
k = cv2.waitKey(1)
if (k & 0xff == ord('q')):
break
c = c + 1
else:
break
# 调用dets里面保存的所有帧的检测信息，进行帧与帧之间的关联
# links = createLinks(dets)
# maxPath(dets, links)
cap.release()
cv2.destroyAllWindows()