def visualization(image, polygons, char_polygons, words, resize_ratio, colors):
draw = ImageDraw.Draw(image, 'RGBA')
for polygon in polygons:
draw.polygon(polygon, fill=None, outline=(0, 255, 0, 255))
for i, char_polygon in enumerate(char_polygons):
for j, polygon in enumerate(char_polygon):
polygon = [int(x * resize_ratio) for x in polygon]
char = words[i][j]
color = colors[char2num(char)]
draw.polygon(polygon, fill=color, outline=color)
def convert(self, mode):
width, height = self.size
if mode == "char_mask":
if not self.use_char_ann:
char_map = -np.ones((height, width))
char_map_weight = np.zeros((self.char_num_classes,))
else:
char_map = np.zeros((height, width))
char_map_weight = np.ones((self.char_num_classes,))
for i, p in enumerate(self.char_boxes):
poly = p.numpy().reshape(4, 2)
poly = shrink_poly(poly, 0.25)
cv2.fillPoly(
char_map, [poly.astype(np.int32)], int(self.char_classes[i])
)
pos_index = np.where(char_map > 0)
pos_num = pos_index[0].size
if pos_num > 0:
pos_weight = 1.0 * (height * width - pos_num) / pos_num
char_map_weight[1:] = pos_weight
return torch.from_numpy(char_map), torch.from_numpy(char_map_weight)
elif mode == "seq_char_mask":
decoder_target = self.char_num_classes * np.ones((32,))
word_target = -np.ones((32,))
if not self.use_char_ann:
char_map = -np.ones((height, width))
char_map_weight = np.zeros((self.char_num_classes,))
for i, char in enumerate(self.word):
if i > 31:
break
decoder_target[i] = char2num(char)
word_target[i] = char2num(char)
end_point = min(max(1, len(self.word)), 31)
word_target[end_point] = self.char_num_classes
else:
char_map = np.zeros((height, width))
char_map_weight = np.ones((self.char_num_classes,))
word_length = 0
for i, p in enumerate(self.char_boxes):
poly = p.numpy().reshape(4, 2)
if i < 32:
decoder_target[i] = int(self.char_classes[i])
word_target[i] = int(self.char_classes[i])
word_length += 1
poly = shrink_poly(poly, 0.25)
cv2.fillPoly(
char_map, [poly.astype(np.int32)], int(self.char_classes[i])
)
end_point = min(max(1, word_length), 31)
word_target[end_point] = self.char_num_classes
pos_index = np.where(char_map > 0)
pos_num = pos_index[0].size
if pos_num > 0:
pos_weight = 1.0 * (height * width - pos_num) / pos_num
char_map_weight[1:] = pos_weight
return (
torch.from_numpy(char_map),
torch.from_numpy(char_map_weight),
torch.from_numpy(decoder_target),
torch.from_numpy(word_target),
)
def convert(self, mode):
width, height = self.size
if mode == "char_mask":
if not self.use_char_ann:
char_map = -np.ones((height, width))
char_map_weight = np.zeros((37,))
else:
char_map = np.zeros((height, width))
char_map_weight = np.ones((37,))
for i, p in enumerate(self.char_boxes):
poly=p.numpy().reshape(4,2)
# x_center = np.mean(poly[:,0], axis = 0).astype(np.int32)
# y_center = np.mean(poly[:,1], axis = 0).astype(np.int32)
poly=shrink_poly(poly,0.25)
cv2.fillPoly(char_map,[poly.astype(np.int32)],int(self.char_classes[i]))
# if is_poly_inbox(poly, height, width) and x_center>=0 and x_center<width and y_center>=0 and y_center<height:
# spoly=shrink_rect(poly,0.25)
# spoly = spoly.astype(np.int32)
# sbox_xmin_shrink = max(0, min(spoly[:,0]))
# sbox_xmax_shrink = min(width - 1, max(spoly[:,0]))
# sbox_ymin_shrink = max(0, min(spoly[:,1]))
# sbox_ymax_shrink = min(height - 1, max(spoly[:,1]))
# ## very small char box
# if sbox_xmax_shrink == sbox_xmin_shrink:
# sbox_xmax_shrink = sbox_xmin_shrink + 1
# if sbox_ymax_shrink == sbox_ymin_shrink:
# sbox_ymax_shrink = sbox_ymin_shrink + 1
# char_map[sbox_ymin_shrink:sbox_ymax_shrink, sbox_xmin_shrink:sbox_xmax_shrink] = int(self.char_classes[i])
pos_index = np.where(char_map > 0)
pos_num = pos_index[0].size
if pos_num > 0:
pos_weight = 1.0 * (height*width - pos_num)/pos_num
char_map_weight[1:] = pos_weight
return torch.from_numpy(char_map), torch.from_numpy(char_map_weight)
elif mode == "seq_char_mask":
decoder_target = (38 - 1) * np.ones((32, ))
word_target = -np.ones((32, ))
if not self.use_char_ann:
char_map = -np.ones((height, width))
char_map_weight = np.zeros((37,))
for i, char in enumerate(self.word):
if i > 31:
break
decoder_target[i] = char2num(char)
word_target[i] = char2num(char)
end_point = min(max(1, len(self.word)), 31)
word_target[end_point] = 37
else:
char_map = np.zeros((height, width))
char_map_weight = np.ones((37,))
word_length = 0
for i, p in enumerate(self.char_boxes):
poly=p.numpy().reshape(4,2)
# x_center = np.mean(poly[:,0], axis = 0).astype(np.int32)
# y_center = np.mean(poly[:,1], axis = 0).astype(np.int32)
# if is_poly_inbox(poly, height, width):
if i < 32:
decoder_target[i] = int(self.char_classes[i])
word_target[i] = int(self.char_classes[i])
word_length += 1
poly=shrink_poly(poly,0.25)
cv2.fillPoly(char_map,[poly.astype(np.int32)],int(self.char_classes[i]))
# if x_center>=0 and x_center<width and y_center>=0 and y_center<height:
# spoly=shrink_rect(poly,0.25)
# spoly = spoly.astype(np.int32)
# sbox_xmin_shrink = max(0, min(spoly[:,0]))
# sbox_xmax_shrink = min(width - 1, max(spoly[:,0]))
# sbox_ymin_shrink = max(0, min(spoly[:,1]))
# sbox_ymax_shrink = min(height - 1, max(spoly[:,1]))
# ## very small char box
# if sbox_xmax_shrink == sbox_xmin_shrink:
# sbox_xmax_shrink = sbox_xmin_shrink + 1
# if sbox_ymax_shrink == sbox_ymin_shrink:
# sbox_ymax_shrink = sbox_ymin_shrink + 1
# char_map[sbox_ymin_shrink:sbox_ymax_shrink, sbox_xmin_shrink:sbox_xmax_shrink] = int(self.char_classes[i])
end_point = min(max(1, word_length), 31)
word_target[end_point] = 37
pos_index = np.where(char_map > 0)
pos_num = pos_index[0].size
if pos_num > 0:
pos_weight = 1.0 * (height*width - pos_num)/pos_num
char_map_weight[1:] = pos_weight
return torch.from_numpy(char_map), torch.from_numpy(char_map_weight), torch.from_numpy(decoder_target), torch.from_numpy(word_target)