def add_shapes(
background: PIL.Image.Image,
shape_img: PIL.Image.Image,
shape_params,
) -> Tuple[List[Tuple[int, int, int, int, int]], PIL.Image.Image]:
"""Paste shapes onto background and return bboxes"""
shape_bboxes: List[Tuple[int, int, int, int, int]] = []
for i, shape_param in enumerate(shape_params):
x = shape_param[-2]
y = shape_param[-1]
x1, y1, x2, y2 = shape_img.getbbox()
bg_at_shape = background.crop((x1 + x, y1 + y, x2 + x, y2 + y))
bg_at_shape.paste(shape_img, (0, 0), shape_img)
background.paste(bg_at_shape, (x, y))
# Slightly expand the bounding box in order to simulate variability with
# the detection boxes. Always make the crop larger than needed because training
# augmentations will only be able to crop down.
dx = random.randint(0, int(0.1 * (x2 - x1)))
dy = random.randint(0, int(0.1 * (y2 - y1)))
x1 -= dx
x2 += dx
y1 -= dy
y2 += dy
background = background.crop((x1 + x, y1 + y, x2 + x, y2 + y))
background = background.filter(ImageFilter.SMOOTH_MORE)
return background.convert("RGB")
def grid(image: PIL.Image.Image, grid_size,
overlap) -> List[List[PIL.Image.Image]]:
the_grid = []
w, h = image.size
for x in range(0, w, grid_size):
row = []
for y in range(0, h, grid_size):
row.append(
image.crop((x - overlap // 2, y - overlap // 2,
x + grid_size + overlap // 2,
y + grid_size + overlap // 2)))
the_grid.append(row)
return the_grid
def strip_image(image: PIL.Image.Image) -> PIL.Image.Image:
"""Remove white and black edges"""
for x in range(image.width):
for y in range(image.height):
r, g, b, _ = image.getpixel((x, y))
if r > 247 and g > 247 and b > 247:
image.putpixel((x, y), (0, 0, 0, 0))
image = image.crop(image.getbbox())
return image
def add_shapes(
background: PIL.Image.Image,
shape_imgs: PIL.Image.Image,
shape_params,
blur_radius: int,
) -> Tuple[List[Tuple[int, int, int, int, int]], PIL.Image.Image]:
"""Paste shapes onto background and return bboxes"""
shape_bboxes: List[Tuple[int, int, int, int, int]] = []
for i, shape_param in enumerate(shape_params):
x = shape_param[-2]
y = shape_param[-1]
shape_img = shape_imgs[i]
shape_img = shape_img.filter(ImageFilter.GaussianBlur(1))
x1, y1, x2, y2 = shape_img.getbbox()
bg_at_shape = background.crop((x1 + x, y1 + y, x2 + x, y2 + y))
bg_at_shape.paste(shape_img, (0, 0), shape_img)
bg_at_shape = bg_at_shape.filter(ImageFilter.SMOOTH_MORE)
background.paste(bg_at_shape, (x, y))
im_w, im_h = background.size
x /= im_w
y /= im_h
w = (x2 - x1) / im_w
h = (y2 - y1) / im_h
shape_bboxes.append((CLASSES.index(shape_param[0]), x, y, w, h))
"""
shape_bboxes.append(
(
CLASSES.index(shape_param[2]),
x + (0.1 * w),
y + (0.1 * h),
0.8 * w,
0.8 * h,
)
)
"""
return shape_bboxes, background.convert("RGB")
def _trim_image(
img: PIL.Image.Image,
border_color: typing.Tuple[int] = None,
) -> PIL.Image.Image:
if border_color is None:
trimmed = img
for xy in itertools.product(range(2), range(2)):
trimmed = _trim_image(
img=trimmed,
border_color=img.getpixel(xy),
)
return trimmed
bg = PIL.Image.new(img.mode, img.size, border_color)
diff = PIL.ImageChops.difference(img, bg)
bbox = diff.getbbox()
if bbox:
return img.crop(bbox)
else:
# found no content
raise ValueError("cannot trim; image was empty")
