def visualize_network_output(output, tr_mask, mode='train'):
vis_dir = os.path.join(cfg.vis_dir, cfg.exp_name + '_' + mode)
if not os.path.exists(vis_dir):
os.mkdir(vis_dir)
b, c, _, _ = output.shape
for i in range(b):
predict = torch.sigmoid(output[i]).data.cpu().numpy()
target = tr_mask[i].cpu().numpy()
shows = list()
for j in range(c):
p = predict[j]
t = target[:, :, j]
tcl_pred = cav.heatmap(
np.array(p / np.max(p) * 255, dtype=np.uint8))
tcl_targ = cav.heatmap(
np.array(t / np.max(t) * 255, dtype=np.uint8))
show = np.concatenate([tcl_pred * 255, tcl_targ * 255], axis=0)
shows.append(show)
show_img = np.concatenate(shows, axis=1)
show = cv2.resize(show_img, (256 * c, 512))
path = os.path.join(vis_dir, '{}.png'.format(i))
cv2.imwrite(path, show)
def visualize_gt(image, contours, tr=None):
image_show = image.copy()
image_show = np.ascontiguousarray(image_show[:, :, ::-1])
image_show = cv2.polylines(image_show, contours, True, (0, 0, 255), 3)
if tr is not None:
tr_map = cav.heatmap(np.array(tr[:, :, -1] / np.max(tr[:, :, -1]) * 255, dtype=np.uint8))
h, w = tr_map.shape[:2]
tr_map = cv2.resize(tr_map, (w*cfg.scale, h*cfg.scale))
image_show = np.concatenate([image_show, np.array(tr_map*255, dtype=np.uint8)], axis=1)
return image_show
else:
return image_show
def visualize_detection(image, contours, tr=None):
image_show = image.copy()
image_show = np.ascontiguousarray(image_show[:, :, ::-1])
cv2.drawContours(image_show, contours, -1, (0, 255, 0), 2)
if tr is not None:
tr_map = cav.heatmap(np.array(tr / np.max(tr) * 255, dtype=np.uint8))
image_show = np.concatenate(
[image_show, np.array(tr_map * 255, dtype=np.uint8)], axis=1)
return image_show
else:
return image_show
def visualize_detection(image, contours, tr=None):
image_show = image.copy()
image_show = np.ascontiguousarray(image_show[:, :, ::-1])
cv2.drawContours(image_show, contours, -1, (0, 255, 0), 3)
if tr is not None:
tr_map = cav.heatmap(np.array(tr / np.max(tr) * 255, dtype=np.uint8))
h,w = tr_map.shape[:2]
tr_map = cv2.resize(tr_map, (w*cfg.scale, h*cfg.scale))
#print(tr_map.shape)
#print(image_show.shape)
image_show = np.concatenate([image_show, np.array(tr_map*255, dtype=np.uint8)], axis=1)
return image_show
else:
return image_show
img = img.transpose(1, 2, 0)
img = ((img * stds + means) * 255).astype(np.uint8)
print(idx, img.shape)
top_map = radius_map[:, :, 0]
bot_map = radius_map[:, :, 1]
print(radius_map.shape)
sin_map, cos_map = regularize_sin_cos(sin_map, cos_map)
ret, labels = cv2.connectedComponents(tcl_mask[:, :,
0].astype(np.uint8),
connectivity=8)
cv2.imshow(
"labels0",
cav.heatmap(np.array(labels * 255 / np.max(labels),
dtype=np.uint8)))
print(np.sum(tcl_mask[:, :, 1]))
t0 = time.time()
for bbox_idx in range(1, ret):
bbox_mask = labels == bbox_idx
text_map = tcl_mask[:, :, 0] * bbox_mask
boxes = bbox_transfor_inv(radius_map,
sin_map,
cos_map,
text_map,
wclip=(2, 8))
# nms
boxes = lanms.merge_quadrangle_n9(boxes.astype('float32'), 0.25)
boxes = boxes[:, :8].reshape((-1, 4, 2)).astype(np.int32)
stds = (0.229, 0.224, 0.225)
transform = Augmentation(
size=640, mean=means, std=stds
)
trainset = Icdar15Text(
data_root='../data/Icdar2015',
is_training=True,
transform=transform
)
# img, train_mask, tr_mask, tcl_mask, radius_map, sin_map, cos_map, meta = trainset[944]
for idx in range(0, len(trainset)):
t0 = time.time()
img, train_mask, tr_mask = trainset[idx]
img, train_mask, tr_mask = map(lambda x: x.cpu().numpy(), (img, train_mask, tr_mask))
img = img.transpose(1, 2, 0)
img = ((img * stds + means) * 255).astype(np.uint8)
print(idx, img.shape)
for i in range(tr_mask.shape[2]):
cv2.imshow("tr_mask_{}".format(i),
cav.heatmap(np.array(tr_mask[:, :, i] * 255 / np.max(tr_mask[:, :, i]), dtype=np.uint8)))
cv2.imshow("train_mask", cav.heatmap(np.array(train_mask * 255 / np.max(train_mask), dtype=np.uint8)))
cv2.imshow('imgs', img)
cv2.waitKey(0)
)
# img, train_mask, tr_mask, tcl_mask, radius_map, sin_map, cos_map, meta = trainset[944]
for idx in range(0, len(trainset)):
t0 = time.time()
img, train_mask, tr_mask, meta = trainset[idx]
# img, train_mask, tr_mask = map(lambda x: x.cpu().numpy(), (img, train_mask, tr_mask))
img = img.transpose(1, 2, 0)
img = ((img * stds + means) * 255).astype(np.uint8)
print(idx, img.shape)
gt_contour = []
for annot, n_annot in zip(meta['annotation'], meta['n_annotation']):
if n_annot.item() > 0:
gt_contour.append(np.array(annot[:n_annot], dtype=np.int))
image_show = cv2.polylines(img, gt_contour, True, (0, 0, 255), 3)
for i in range(tr_mask.shape[2]):
heatmap = cav.heatmap(np.array(tr_mask[:, :, i] * 255 / np.max(tr_mask[:, :, i]), dtype=np.uint8))
cv2.imshow("tr_mask_{}".format(i),heatmap)
cv2.imwrite("{}.png".format(i), heatmap*255)
mask = tr_mask[:, :, 0]
# from scipy import ndimage as ndimg
# dmp = ndimg.distance_transform_edt(mask) # distance transform
cv2.imwrite("mask1.png".format(i), np.array(mask*255, dtype=np.uint8))
cv2.imwrite("mask0.png".format(i), np.array((mask>0) * 255, dtype=np.uint8))
cv2.imshow('imgs', image_show)
cv2.imwrite("imgs.png".format(i), image_show)
cv2.waitKey(0)