def evaluate_textboxes(gt_boxes, boxes):
"""
Evaluates the mean intersection-over-union between GT textboxes and the given ones.
Parameters
----------
gt_boxes : list of textboxes for each image, as described in W2 slides
boxes : list of textboxes for each image, as described in W2 slides
Returns
-------
mean_iou: mean intersection-over-union
"""
assert len(gt_boxes) == len(boxes)
iou = 0
# compute IOU per image
for i in range(len(boxes)):
if len(boxes[i]) == 0 or len(gt_boxes[i]) == 0:
continue
max_dim = np.max(np.max(boxes[i]))
shape = (max_dim, max_dim)
# We compute the IOU by generating both masks with all given textboxes highlighted.
gt_mask, mask = generate_text_mask(shape,
gt_boxes[i]), generate_text_mask(
shape, boxes[i])
iou += compute_iou(gt_mask, mask)
return iou / len(boxes)
def _extract_image_and_mask_from_path(img_path, mask_path, textboxes, multiple):
img = path2img(img_path)
if mask_path is None:
text_mask = 1 - generate_text_mask(img.shape[:2], textboxes) # returns all 1's if textboxes == None
return [(img, text_mask), ]
mask = path2mask(mask_path)
text_mask = 1 - generate_text_mask(mask.shape, textboxes)
if multiple:
masks = extract_paintings_from_mask(mask)
if masks is None or len(masks) == 0:
return [(img, np.ones(img.shape[:2]).astype(np.uint8)), ]
return [(img, ((text_mask != 0) & (m != 0)).astype(np.uint8)) for m in masks]
return [(img, ((text_mask != 0) & (mask != 0)).astype(np.uint8)), ]
def _extract_features_from_path(_path, textboxes):
img = path2img(_path)
text_mask = 1 - generate_text_mask(
img.shape[:2], textboxes) # returns all 1's if textboxes == None
return [
extract_features_func(img, mask=text_mask),
]
def _extract_features_from_path_mask(_path, _path_mask, textboxes):
img, mask = path2img(_path), path2mask(_path_mask)
text_mask = 1 - generate_text_mask(mask.shape, textboxes)
if multiple:
masks = extract_paintings_from_mask(mask)
#print(f"{_path_mask} - {len(masks)}")
return [
extract_features_func(img,
mask=((text_mask != 0) &
(m != 0)).astype(np.uint8))
for m in masks
]
return [
extract_features_func(img,
mask=((text_mask != 0) & (mask != 0)).astype(
np.uint8)),
]
