def overlap_expand(math_regions):
print('Number of math regions ', len(math_regions))
if type(math_regions) != type([]):
math_regions = math_regions.tolist()
obsolete = []
for i in range(len(math_regions)):
for j in range(i + 1, len(math_regions)):
# print(i,j)
if box_utils.intersects(math_regions[i], math_regions[j]):
math_regions[i][0] = min(math_regions[i][0],
math_regions[j][0])
math_regions[i][1] = min(math_regions[i][1],
math_regions[j][1])
math_regions[i][2] = max(math_regions[i][2],
math_regions[j][2])
math_regions[i][3] = max(math_regions[i][3],
math_regions[j][3])
obsolete.append(j)
math_regions = [i for j, i in enumerate(math_regions) if j not in obsolete]
return math_regions
def generate_metadata(self):
for id in self.ids:
math_patches = []
height, width, channels = self.images[id[1]].shape
current_page_boxes = self.math_ground_truth[id[1]]
n_horizontal = np.ceil(width / self.window) # 4
n_vertical = np.ceil(height / self.window) # 5
h = np.arange(0, n_horizontal - 1 + self.stride, self.stride)
v = np.arange(0, n_vertical - 1 + self.stride, self.stride)
crop_size = self.window
if (self.split == 'train' or self.split == 'validate') and self.is_math[id[1]]:
for i in h:
for j in v:
x_l = int(np.round(crop_size * i))
x_h = x_l + self.window
y_l = int(np.round(crop_size * j))
y_h = y_l + self.window
# left, top, right, bottom
image_box = [x_l, y_l, x_h, y_h]
current_page_boxes = copy.deepcopy(self.math_ground_truth[id[1]])
# if math intersects only consider the region which
# is part of the current bounding box
for box in current_page_boxes:
if box_utils.intersects(image_box, box):
# left, top, right, bottom
# y increases downwards
# crop the boxes to fit into image region
box[0] = max(x_l, box[0])
box[1] = max(y_l, box[1])
box[2] = min(x_h, box[2])
box[3] = min(y_h, box[3])
# # Translate to origin
box[0] = box[0] - x_l
box[2] = box[2] - x_l
box[1] = box[1] - y_l
box[3] = box[3] - y_l
if feature_extractor.width(box) > 0 and feature_extractor.height(box) > 0:
self.metadata.append([id[1], x_l, y_l])
break
elif self.split=='test':
for i in h:
for j in v:
x_l = int(np.round(crop_size * i))
y_l = int(np.round(crop_size * j))
self.metadata.append([id[1], x_l, y_l])
def gen_targets(self, index):
metadata = self.metadata[index]
x_l = metadata[1]
y_l = metadata[2]
x_h = x_l + self.window
y_h = y_l + self.window
current_page_boxes = copy.deepcopy(self.math_ground_truth[metadata[0]])
targets = []
image_box = [x_l, y_l, x_h, y_h]
# if math intersects only consider the region which
# is part of the current bounding box
for box in current_page_boxes:
if box_utils.intersects(image_box, box):
# left, top, right, bottom
# y increases downwards
# crop the boxes to fit into image region
box[0] = max(x_l, box[0])
box[1] = max(y_l, box[1])
box[2] = min(x_h, box[2])
box[3] = min(y_h, box[3])
# # Translate to origin
box[0] = box[0] - x_l
box[2] = box[2] - x_l
box[1] = box[1] - y_l
box[3] = box[3] - y_l
if feature_extractor.width(
box) > 0 and feature_extractor.height(box) > 0:
targets.append(box)
# It is done only for testing, where we do not care about targets
# This avoids IndexError: too many indices for array
# TODO: refactor in future
if len(targets) == 0:
targets = [[-1, -1, -1, -1]]
return targets