def translate(image, annotations, prob=0.5, border_value=(128, 128, 128)):
assert 'bboxes' in annotations, 'annotations should contain bboxes even if it is empty'
random_prob = np.random.uniform()
if random_prob < (1 - prob):
return image, annotations
h, w = image.shape[:2]
bboxes = annotations['bboxes']
if bboxes.shape[0] != 0:
min_x1, min_y1 = np.min(bboxes, axis=0)[:2].astype(np.int32)
max_x2, max_y2 = np.max(bboxes, axis=0)[2:].astype(np.int32)
translation_matrix = translation_xy(min=(min(-(min_x1 // 2), 0), min(-(min_y1 // 2), 0)),
max=(max((w - 1 - max_x2) // 2, 1), max((h - 1 - max_y2) // 2, 1)),
prob=1.)
else:
translation_matrix = translation_xy(min=(min(-w // 8, 0), min(-h // 8, 0)),
max=(max(w // 8, 1), max(h // 8, 1)))
translation_matrix = change_transform_origin(translation_matrix, (w / 2, h / 2))
image = cv2.warpAffine(
image,
translation_matrix[:2, :],
dsize=(w, h),
flags=cv2.INTER_CUBIC,
borderMode=cv2.BORDER_CONSTANT,
borderValue=border_value,
)
if bboxes.shape[0] != 0:
new_bboxes = []
for bbox in bboxes:
x1, y1, x2, y2 = bbox
points = translation_matrix.dot([
[x1, x2, x1, x2],
[y1, y2, y2, y1],
[1, 1, 1, 1],
])
min_x, min_y = np.min(points, axis=1)[:2]
max_x, max_y = np.max(points, axis=1)[:2]
new_bboxes.append([min_x, min_y, max_x, max_y])
annotations['bboxes'] = np.array(new_bboxes).astype(np.float32)
if 'quadrangles' in annotations and annotations['quadrangles'].shape[0] != 0:
quadrangles = annotations['quadrangles']
translated_quadrangles = []
for quadrangle in quadrangles:
quadrangle = np.concatenate([quadrangle, np.ones((4, 1))], axis=-1)
translated_quadrangle = translation_matrix.dot(quadrangle.T).T[:, :2]
quadrangle = reorder_vertexes(translated_quadrangle)
translated_quadrangles.append(quadrangle)
quadrangles = np.stack(translated_quadrangles)
annotations['quadrangles'] = quadrangles
xmin = np.min(quadrangles, axis=1)[:, 0]
ymin = np.min(quadrangles, axis=1)[:, 1]
xmax = np.max(quadrangles, axis=1)[:, 0]
ymax = np.max(quadrangles, axis=1)[:, 1]
bboxes = np.stack([xmin, ymin, xmax, ymax], axis=1)
annotations['bboxes'] = bboxes
return image, annotations
def rotate(image, annotations, prob=0.5, border_value=(128, 128, 128)):
assert 'bboxes' in annotations, 'annotations should contain bboxes even if it is empty'
random_prob = np.random.uniform()
if random_prob < (1 - prob):
return image, annotations
rotate_degree = np.random.uniform(low=-45, high=45)
h, w = image.shape[:2]
# Compute the rotation matrix.
M = cv2.getRotationMatrix2D(center=(w / 2, h / 2),
angle=rotate_degree,
scale=1)
# Get the sine and cosine from the rotation matrix.
abs_cos_angle = np.abs(M[0, 0])
abs_sin_angle = np.abs(M[0, 1])
# Compute the new bounding dimensions of the image.
new_w = int(h * abs_sin_angle + w * abs_cos_angle)
new_h = int(h * abs_cos_angle + w * abs_sin_angle)
# Adjust the rotation matrix to take into account the translation.
M[0, 2] += new_w // 2 - w // 2
M[1, 2] += new_h // 2 - h // 2
# Rotate the image.
image = cv2.warpAffine(image,
M=M,
dsize=(new_w, new_h),
flags=cv2.INTER_CUBIC,
borderMode=cv2.BORDER_CONSTANT,
borderValue=border_value)
bboxes = annotations['bboxes']
if bboxes.shape[0] != 0:
new_bboxes = []
for bbox in bboxes:
x1, y1, x2, y2 = bbox
points = M.dot([
[x1, x2, x1, x2],
[y1, y2, y2, y1],
[1, 1, 1, 1],
])
# Extract the min and max corners again.
min_xy = np.sort(points, axis=1)[:, :2]
min_x = np.mean(min_xy[0])
min_y = np.mean(min_xy[1])
max_xy = np.sort(points, axis=1)[:, 2:]
max_x = np.mean(max_xy[0])
max_y = np.mean(max_xy[1])
new_bboxes.append([min_x, min_y, max_x, max_y])
annotations['bboxes'] = np.array(new_bboxes, dtype=np.float32)
if 'quadrangles' in annotations and annotations['quadrangles'].shape[
0] != 0:
quadrangles = annotations['quadrangles']
rotated_quadrangles = []
for quadrangle in quadrangles:
quadrangle = np.concatenate(
[quadrangle, np.ones((4, 1))], axis=-1)
rotated_quadrangle = M.dot(quadrangle.T).T[:, :2]
quadrangle = reorder_vertexes(rotated_quadrangle)
rotated_quadrangles.append(quadrangle)
quadrangles = np.stack(rotated_quadrangles)
annotations['quadrangles'] = quadrangles
xmin = np.min(quadrangles, axis=1)[:, 0]
ymin = np.min(quadrangles, axis=1)[:, 1]
xmax = np.max(quadrangles, axis=1)[:, 0]
ymax = np.max(quadrangles, axis=1)[:, 1]
bboxes = np.stack([xmin, ymin, xmax, ymax], axis=1)
annotations['bboxes'] = bboxes
return image, annotations