max_im_shrink = (
0x7fffffff / 577.0 /
(img.shape[0] * img.shape[1]))**0.5 # the max size of input image
shrink = max_im_shrink if max_im_shrink < 1 else 1
# print('max_im_shrink:', max_im_shrink)
det0 = detect_face(img)
det1 = detect_face(img, flip=True)
# det = np.row_stack((det0, det1))
det2 = im_det_ms_pyramid(img, max_im_shrink)
# merge all test results via bounding box voting
det = np.row_stack((det0, det1, det2))
keep_index = np.where(
np.minimum(det[:, 2] - det[:, 0], det[:, 3] -
det[:, 1]) >= 2)[0] # >= 3
dets = det[keep_index, :]
dets = bbox_process.soft_bbox_vote(dets, thre=soft_box_thre, score=score)
# keep = nms(dets, 0.4, usegpu=True, gpu_id=0)
# dets = det[keep, :]
keep_index = np.where((dets[:, 2] - dets[:, 0] + 1) *
(dets[:, 3] - dets[:, 1] + 1) >= 2**2)[0]
dets = dets[keep_index, :]
with open(txtpath, 'w') as file:
file.write('{:s}\n'.format(filename))
file.write('{:d}\n'.format(len(dets)))
for line in dets:
file.write('{:.0f} {:.0f} {:.0f} {:.0f} {:.3f}\n'.format(
line[0], line[1], line[2] - line[0] + 1, line[3] - line[1] + 1,
line[4]))
elif st[i] == 2.0:
index = np.where(
np.minimum(det_temp[:, 2] - det_temp[:, 0] + 1, det_temp[:, 3] - det_temp[:, 1] + 1) < 48)[0]
det_temp = det_temp[index, :]
elif st[i] == 2.25:
index = np.where(
np.minimum(det_temp[:, 2] - det_temp[:, 0] + 1, det_temp[:, 3] - det_temp[:, 1] + 1) < 32)[0]
det_temp = det_temp[index, :]
det_s = np.row_stack((det_s, det_temp))
return det_s
max_im_shrink = (0x7fffffff / 577.0 / (img.shape[0] * img.shape[1])) ** 0.5 # the max size of input image
shrink = max_im_shrink if max_im_shrink < 1 else 1
det0 = detect_face(img)
det1 = detect_face(img, flip=True)
det2 = im_det_ms_pyramid(img, max_im_shrink)
# merge all test results via bounding box voting
det = np.row_stack((det0, det1, det2))
keep_index = np.where(np.minimum(det[:, 2] - det[:, 0], det[:, 3] - det[:, 1]) >= 3)[0]
det = det[keep_index, :]
dets = bbox_process.soft_bbox_vote(det, thre=0.4)
keep_index = np.where((dets[:, 2] - dets[:, 0] + 1) * (dets[:, 3] - dets[:, 1] + 1) >= 6 ** 2)[0]
dets = dets[keep_index, :]
with open(txtpath, 'w') as f:
f.write('{:s}\n'.format(filename))
f.write('{:d}\n'.format(len(dets)))
for line in dets:
f.write('{:.0f} {:.0f} {:.0f} {:.0f} {:.3f}\n'.
format(line[0], line[1], line[2] - line[0] + 1, line[3] - line[1] + 1, line[4]))
print time.time() - start_time
def detect_main(model, val_data, C):
C.size_test = [0, 0]
num_imgs = len(val_data)
print 'num of val samples: {}'.format(num_imgs)
meta = {}
meta['filepath'] = []
meta['true_box'] = []
meta['pred_box'] = []
progbar = generic_utils.Progbar(num_imgs)
start_time = time.time()
for f in range(num_imgs):
filepath = val_data[f]['filepath']
# filename = filepath.split('/')[-1].split('.')[0]
img = cv2.imread(filepath)
max_im_shrink = (
0x7fffffff / 577.0 /
(img.shape[0] * img.shape[1]))**0.5 # the max size of input image
det0 = detect_face(model, img, C)
det1 = detect_face(model, img, C, flip=True)
det2 = im_det_ms_pyramid(model, img, C, max_im_shrink)
det = np.row_stack((det0, det1, det2))
keep_index = np.where(
np.minimum(det[:, 2] - det[:, 0], det[:, 3] -
det[:, 1]) >= 2)[0] # >= 3
dets = det[keep_index, :]
dets = bbox_process.soft_bbox_vote(dets,
thre=soft_box_thre,
score=score)
keep_index = np.where((dets[:, 2] - dets[:, 0] + 1) *
(dets[:, 3] - dets[:, 1] + 1) >= 2**2)[0]
dets = dets[keep_index, :]
gt_bboxes = val_data[f]['bboxes']
meta['filepath'].append(filepath)
meta['true_box'].append(dets)
meta['pred_box'].append(gt_bboxes)
if f % 10 == 0:
progbar.update(f)
best_mae = 1e8
best_mse = 1e8
best_thr = 0.4
mae_dict = {}
mse_dict = {}
for i in range(30, 51):
thr_i = i * 0.01
pred_counts = []
true_counts = []
for true_box, pred_box in zip(meta['true_box'], meta['pred_box']):
true_box = true_box[true_box[:, 4] > thr_i]
pred_count = true_box.shape[0]
true_count = pred_box.shape[0]
pred_counts.append(pred_count)
true_counts.append(true_count)
pred_counts = np.array(pred_counts, dtype='float')
true_counts = np.array(true_counts, dtype='float')
mae = np.mean(np.abs(true_counts - pred_counts))
mse = np.sqrt(np.mean((true_counts - pred_counts)**2))
mae_dict['{:0.2f}'.format(thr_i)] = float(mae)
mse_dict['{:0.2f}'.format(thr_i)] = float(mse)
if mae <= best_mae:
best_mae = float(mae)
best_mse = float(mse)
best_thr = float(thr_i)
print time.time() - start_time
return best_mae, best_mse, best_thr, mae_dict, mse_dict
