res_gt = '{0}/gt_{1}'.format(args.images_dir, res_gt)
if not os.path.exists(res_gt):
print('missing! {0}'.format(res_gt))
gt_rect, gt_txts = [], []
#continue
gt_rect, gt_txts = load_gt(res_gt)
print(img_name)
img = cv2.imread(img_name)
#font = cv2.FONT_HERSHEY_SIMPLEX
#cv2.putText(img,'cs',(10,img.shape[0] -40), font, 0.8,(255,255,255),2,cv2.LINE_AA)
im_resized, (ratio_h, ratio_w) = resize_image(
img, max_size=1848 * 1024,
scale_up=True) #1348*1024 #1848*1024
#im_resized = im_resized[:, :, ::-1]
images = np.asarray([im_resized], dtype=np.float)
images /= 128
images -= 1
im_data = net_utils.np_to_variable(images,
is_cuda=args.cuda).permute(
0, 3, 1, 2)
[iou_pred, iou_pred1], rboxs, angle_pred, features = net(im_data)
iou = iou_pred.data.cpu()[0].numpy()
iou = iou.squeeze(0)
iou_pred1 = iou_pred1.data.cpu()[0].numpy()
iou_pred1 = iou_pred1.squeeze(0)
def evaluate_e2e_crnn(root,
net,
norm_height=48,
name_model='E2E',
normalize=False,
save=False,
cuda=True,
save_dir='eval'):
#Decription : evaluate model E2E
net = net.eval()
# if cuda:
# print('Using cuda ...')
# net = net.to(device)
images = glob.glob(os.path.join(root, '*.jpg'))
png = glob.glob(os.path.join(root, '*.png'))
images.extend(png)
png = glob.glob(os.path.join(root, '*.JPG'))
images.extend(png)
imagess = np.asarray(images)
tp_all = 0
gt_all = 0
tp_e2e_all = 0
gt_e2e_all = 0
tp_e2e_ed1_all = 0
detecitons_all = 0
eval_text_length = 2
segm_thresh = 0.5
min_height = 8
idx = 0
if not os.path.exists(save_dir):
os.mkdir(save_dir)
note_path = os.path.join(save_dir, 'note_eval.txt')
note_file = open(note_path, 'a')
with torch.no_grad():
index = np.arange(0, imagess.shape[0])
# np.random.shuffle(index)
for i in index:
img_name = imagess[i]
base_nam = os.path.basename(img_name)
#
# if args.evaluate == 1:
res_gt = base_nam.replace(".jpg", '.txt').replace(".png", '.txt')
res_gt = '{0}/gt_{1}'.format(root, res_gt)
if not os.path.exists(res_gt):
res_gt = base_nam.replace(".jpg", '.txt').replace("_", "")
res_gt = '{0}/gt_{1}'.format(root, res_gt)
if not os.path.exists(res_gt):
print('missing! {0}'.format(res_gt))
gt_rect, gt_txts = [], []
# continue
gt_rect, gt_txts = load_gt(res_gt)
# print(img_name)
img = cv2.imread(img_name)
im_resized, _ = resize_image(
img, max_size=1848 * 1024,
scale_up=False) # 1348*1024 #1848*1024
images = np.asarray([im_resized], dtype=np.float)
if normalize:
images /= 128
images -= 1
im_data = net_utils.np_to_variable(images, is_cuda=cuda).permute(
0, 3, 1, 2)
[iou_pred, iou_pred1], rboxs, angle_pred, features = net(im_data)
iou = iou_pred.data.cpu()[0].numpy()
iou = iou.squeeze(0)
rbox = rboxs[0].data.cpu()[0].numpy()
rbox = rbox.swapaxes(0, 1)
rbox = rbox.swapaxes(1, 2)
detections = get_boxes(iou, rbox,
angle_pred[0].data.cpu()[0].numpy(),
segm_thresh)
im_scalex = im_resized.shape[1] / img.shape[1]
im_scaley = im_resized.shape[0] / img.shape[0]
detetcions_out = []
detectionso = np.copy(detections)
if len(detections) > 0:
detections[:, 0] /= im_scalex
detections[:, 2] /= im_scalex
detections[:, 4] /= im_scalex
detections[:, 6] /= im_scalex
detections[:, 1] /= im_scaley
detections[:, 3] /= im_scaley
detections[:, 5] /= im_scaley
detections[:, 7] /= im_scaley
for bid, box in enumerate(detections):
boxo = detectionso[bid]
# score = boxo[8]
boxr = boxo[0:8].reshape(-1, 2)
# box_area = area(boxr.reshape(8))
# conf_factor = score / box_area
center = (boxr[0, :] + boxr[1, :] + boxr[2, :] +
boxr[3, :]) / 4
dw = boxr[2, :] - boxr[1, :]
dw2 = boxr[0, :] - boxr[3, :]
dh = boxr[1, :] - boxr[0, :]
dh2 = boxr[3, :] - boxr[2, :]
h = math.sqrt(dh[0] * dh[0] + dh[1] * dh[1]) + 1
h2 = math.sqrt(dh2[0] * dh2[0] + dh2[1] * dh2[1]) + 1
h = (h + h2) / 2
w = math.sqrt(dw[0] * dw[0] + dw[1] * dw[1])
w2 = math.sqrt(dw2[0] * dw2[0] + dw2[1] * dw2[1])
w = (w + w2) / 2
if ((h - 1) / im_scaley) < min_height:
continue
input_W = im_data.size(3)
input_H = im_data.size(2)
target_h = norm_height
scale = target_h / h
target_gw = int(w * scale + target_h / 4)
target_gw = max(8, int(round(target_gw / 8)) * 8)
xc = center[0]
yc = center[1]
w2 = w
h2 = h
angle = math.atan2((boxr[2][1] - boxr[1][1]),
boxr[2][0] - boxr[1][0])
angle2 = math.atan2((boxr[3][1] - boxr[0][1]),
boxr[3][0] - boxr[0][0])
angle = (angle + angle2) / 2
# show pooled image in image layer
scalex = (w2 + h2 / 4) / input_W
scaley = h2 / input_H
th11 = scalex * math.cos(angle)
th12 = -math.sin(angle) * scaley * input_H / input_W
th13 = (2 * xc - input_W - 1) / (input_W - 1)
th21 = math.sin(angle) * scalex * input_W / input_H
th22 = scaley * math.cos(angle)
th23 = (2 * yc - input_H - 1) / (input_H - 1)
t = np.asarray([th11, th12, th13, th21, th22, th23],
dtype=np.float)
t = torch.from_numpy(t).type(torch.FloatTensor)
t = t.to(device)
theta = t.view(-1, 2, 3)
grid = F.affine_grid(
theta, torch.Size((1, 3, int(target_h), int(target_gw))))
x = F.grid_sample(im_data, grid)
# features = net.forward_features(x)
# labels_pred = net.forward_ocr(features)
labels_pred = net.forward_ocr(x)
labels_pred = labels_pred.permute(1, 2, 0)
ctc_f = labels_pred.data.cpu().numpy()
ctc_f = ctc_f.swapaxes(1, 2)
labels = ctc_f.argmax(2)
conf = np.mean(np.exp(ctc_f.max(2)[labels > 3]))
if conf < 0.02:
continue
det_text, conf2, dec_s, word_splits = print_seq_ext(
labels[0, :], codec)
det_text = det_text.strip()
if conf < 0.01 and len(det_text) == 3:
continue
if len(det_text) > 0:
dtxt = det_text.strip()
if len(dtxt) >= eval_text_length:
# print('{0} - {1}'.format(dtxt, conf_factor))
boxw = np.copy(boxr)
boxw[:, 1] /= im_scaley
boxw[:, 0] /= im_scalex
boxw = boxw.reshape(8)
detetcions_out.append([boxw, dtxt])
pix = img
# if args.evaluate == 1:
tp, tp_e2e, gt_e2e, tp_e2e_ed1, detection_to_gt, pixx = evaluate_image(
pix,
detetcions_out,
gt_rect,
gt_txts,
eval_text_length=eval_text_length)
tp_all += tp
gt_all += len(gt_txts)
tp_e2e_all += tp_e2e
gt_e2e_all += gt_e2e
tp_e2e_ed1_all += tp_e2e_ed1
detecitons_all += len(detetcions_out)
# print(gt_all)
if save:
cv2.imwrite('{0}/{1}'.format(save_dir, base_nam), pixx)
# print(" E2E recall tp_e2e:{0:.3f} / tp:{1:.3f} / e1:{2:.3f}, precision: {3:.3f}".format(
# tp_e2e_all / float(max(1, gt_e2e_all)),
# tp_all / float(max(1, gt_e2e_all)),
# tp_e2e_ed1_all / float(max(1, gt_e2e_all)),
# tp_all / float(max(1, detecitons_all))))
note_file.write(
'Model{4}---E2E recall tp_e2e:{0:.3f} / tp:{1:.3f} / e1:{2:.3f}, precision: {3:.3f} \n'
.format(tp_e2e_all / float(max(1, gt_e2e_all)),
tp_all / float(max(1, gt_e2e_all)),
tp_e2e_ed1_all / float(max(1, gt_e2e_all)),
tp_all / float(max(1, detecitons_all)), name_model))
note_file.close()
return (tp_e2e_all / float(max(1, gt_e2e_all)),
tp_all / float(max(1, gt_e2e_all)),
tp_e2e_ed1_all / float(max(1, gt_e2e_all)),
tp_all / float(max(1, detecitons_all)))