def get_relative_box_position(anchor, box):
anchor_vec = Vec2(anchor.left, anchor.top)
return BBox(
box.top_left - anchor_vec,
box.top_right - anchor_vec,
box.bottom_right - anchor_vec,
box.bottom_left - anchor_vec,
)
def get_bboxes(bboxes):
if len(bboxes.shape) == 2:
bboxes = bboxes[..., None]
bboxes = bboxes.transpose(2, 1, 0)
boxes = []
aabbs = []
for box in bboxes:
box = BBox._make([Vec2._make(b).to_int() for b in box])
boxes.append(box)
aabbs.append(box.to_aabb())
return boxes, aabbs
def crop_chars(self, image, i):
image_size = Size(*image.shape[-2:])
image_bbox = BBox.from_axis_aligned_bbox(
AxisAlignedBBox(0, 0, image_size.width, image_size.height))
transform_params = []
crop_vector = Vec2(x=image_bbox.top_right.x - image_bbox.top_left.x,
y=image_bbox.top_right.y - image_bbox.top_left.y)
start_points = [
(crop_vector * (i / self.num_words_per_image)).to_int()
for i in range(self.num_words_per_image)
]
crop_bbox_width = crop_vector * (1 / (self.num_words_per_image / 4)
) # divide by four for 75% overlap
box_overlap = Vec2(
(start_points[-1].x + crop_bbox_width.x - image_size.width) // 2,
0)
start_points = [point - box_overlap for point in start_points]
with cuda.get_device_from_id(self.device_id):
for start_point in start_points:
top_left = image_bbox.top_left + start_point
bottom_left = image_bbox.bottom_left + start_point
top_right = image_bbox.top_left + start_point + crop_bbox_width
bottom_right = image_bbox.bottom_left + start_point + crop_bbox_width
crop_bbox = BBox(top_left, top_right, bottom_right,
bottom_left)
transform_params.append(
crop_bbox.get_affine_transform_params(Vec2(
x=image_size.width, y=image_size.height),
xp=self.xp))
transform_params = self.xp.stack(transform_params, axis=0)
return self.apply_transform_params(image, transform_params)