def get_annot_from_image_data(self, image_data: ImageData) -> Dict:
image_data = self.filter_image_data(image_data)
image = image_data.open_image()
height, width, _ = image.shape
annot = {
"description":
"",
"tags": [],
"size": {
"height": height,
"width": width
},
"objects": [{
"description":
"",
"geometryType":
"rectangle",
"tags": [{
"name": str(bbox_data.label),
"value": None,
}],
"classTitle":
"bbox",
"points": {
"exterior": [[int(bbox_data.xmin),
int(bbox_data.ymin)],
[int(bbox_data.xmax),
int(bbox_data.ymax)]],
"interior": []
}
} for bbox_data in image_data.bboxes_data]
}
return annot
def apply_perspective_transform_to_image_data(
image_data: ImageData, perspective_matrix: np.ndarray,
result_width: int, result_height: int,
allow_negative_and_large_coords: bool,
remove_bad_coords: bool) -> ImageData:
image = image_data.open_image()
image = cv2.warpPerspective(image, perspective_matrix,
(result_width, result_height))
image_data = copy.deepcopy(image_data)
image_data.keypoints = apply_perspective_transform_to_points(
image_data.keypoints, perspective_matrix, result_width, result_height,
allow_negative_and_large_coords, remove_bad_coords)
image_data.bboxes_data = [
_apply_perspective_transform_to_bbox_data(
bbox_data, perspective_matrix, result_width, result_height,
allow_negative_and_large_coords, remove_bad_coords)
for bbox_data in image_data.bboxes_data
]
image_data.bboxes_data = [
bbox_data for bbox_data in image_data.bboxes_data
if bbox_data is not None
]
image_data.image_path = None
image_data.image = image
return image_data
def thumbnail_image_data(image_data: ImageData,
size: Tuple[int, int],
resample: Optional[int] = None) -> ImageData:
image = image_data.open_image()
new_width, new_height = get_thumbnail_resize(Image.fromarray(image), size)
return resize_image_data(image_data, (new_width, new_height),
resample=resample)
def convert_image_data_to_polygon_label(
image_data: ImageData,
from_name: str,
polygonlabels: str,
) -> Dict:
if image_data.image_path is not None:
im_width, im_height = imagesize.get(image_data.image_path)
else:
im_height, im_width, _ = image_data.open_image().shape
rectangle_labels = []
for bbox_data in image_data.bboxes_data:
rectangle_labels.append({
"original_width": im_width,
"original_height": im_height,
"image_rotation": 0,
"value": {
"points": [[x * 100 / im_width, y * 100 / im_height]
for x, y in bbox_data.keypoints],
"polygonlabels": [polygonlabels]
},
"from_name": from_name,
"to_name": "image",
"type": "polygonlabels"
})
return rectangle_labels
def convert_image_data_to_rectangle_labels(
image_data: ImageData,
from_name: str,
to_name: str,
) -> Dict:
if image_data.image_path is not None:
im_width, im_height = imagesize.get(image_data.image_path)
else:
im_height, im_width, _ = image_data.open_image().shape
rectangle_labels = []
for bbox_data in image_data.bboxes_data:
rectangle_labels.append({
"original_width": im_width,
"original_height": im_height,
"image_rotation": 0,
"value": {
"x": bbox_data.xmin / im_width * 100,
"y": bbox_data.ymin / im_height * 100,
"width": (bbox_data.xmax - bbox_data.xmin) / im_width * 100,
"height": (bbox_data.ymax - bbox_data.ymin) / im_height * 100,
"rotation": 0,
"rectanglelabels": [bbox_data.label]
},
"from_name": from_name,
"to_name": to_name,
"type": "rectanglelabels"
})
return rectangle_labels
def convert_image_data_to_keypoint_label(
image_data: ImageData,
from_name: str,
keypointlabels: str,
) -> Dict:
if image_data.image_path is not None:
im_width, im_height = imagesize.get(image_data.image_path)
else:
im_height, im_width, _ = image_data.open_image().shape
rectangle_labels = []
for bbox_data in image_data.bboxes_data:
for keypoint in bbox_data.keypoints:
x, y = keypoint[0], keypoint[1]
rectangle_labels.append({
"original_width": im_width,
"original_height": im_height,
"image_rotation": 0,
"value": {
"x": x * 100 / im_width,
"y": y * 100 / im_height,
"width": 0.55,
"keypointlabels": [keypointlabels]
},
"from_name": from_name,
"to_name": "image",
"type": "keypointlabels"
})
return rectangle_labels
def rotate_image_data(image_data: ImageData,
angle: float,
border_mode: Optional[int] = None,
border_value: Tuple[int, int, int] = None):
if abs(angle) <= 1e-6:
return image_data
image = image_data.open_image()
height, width, _ = image.shape
image_center = width // 2, height // 2
angle_to_factor = {0: 0, 90: 1, 180: 2, 270: 3}
angle = angle % 360
rotated_image_data = copy.deepcopy(image_data)
if angle in angle_to_factor:
factor = angle_to_factor[angle]
rotated_image = np.rot90(image, factor)
rotated_image_data.keypoints = rotate_keypoints90(
image_data.keypoints, factor, width, height)
rotated_image_data.bboxes_data = [
_rotate_bbox_data90(bbox_data, factor, width, height)
for bbox_data in rotated_image_data.bboxes_data
]
else:
# grab the rotation matrix
rotation_mat = cv2.getRotationMatrix2D(image_center, angle, 1.)
# compute the new bounding dimensions of the image
abs_cos = abs(rotation_mat[0, 0])
abs_sin = abs(rotation_mat[0, 1])
bound_w = int(height * abs_sin + width * abs_cos)
bound_h = int(height * abs_cos + width * abs_sin)
# adjust the rotation matrix to take into account translation
rotation_mat[0, 2] += bound_w / 2 - image_center[0]
rotation_mat[1, 2] += bound_h / 2 - image_center[1]
rotated_image = cv2.warpAffine(image,
rotation_mat, (bound_w, bound_h),
borderMode=border_mode,
borderValue=border_value)
new_height, new_width, _ = rotated_image.shape
rotated_image_data = copy.deepcopy(image_data)
rotated_image_data.keypoints = rotate_keypoints(
image_data.keypoints, rotation_mat, new_height, new_width)
keypoints = []
for (x, y) in rotated_image_data.keypoints:
x = max(0, min(x, new_width - 1))
y = max(0, min(y, new_height - 1))
keypoints.append([x, y])
rotated_image_data.keypoints = np.array(keypoints).reshape(-1, 2)
rotated_image_data.bboxes_data = [
_rotate_bbox_data(bbox_data, rotation_mat, new_height, new_width)
for bbox_data in rotated_image_data.bboxes_data
]
rotated_image_data.image_path = None # It applies to all bboxes_data inside
rotated_image_data.image = rotated_image
return rotated_image_data
def resize_image_data(image_data: ImageData,
size: Tuple[int, int],
resample: Optional[int] = None) -> ImageData:
image_data = copy.deepcopy(image_data)
image = image_data.open_image()
old_height, old_width, _ = image.shape
image = Image.fromarray(image)
image = image.resize(size, resample=resample)
image = np.array(image)
new_height, new_width, _ = image.shape
def resize_coords(bbox_data: BboxData):
bbox_data.xmin = max(
0, min(int(bbox_data.xmin * (new_width / old_width)),
new_width - 1))
bbox_data.ymin = max(
0,
min(int(bbox_data.ymin * (new_height / old_height)),
new_height - 1))
bbox_data.xmax = max(
0, min(int(bbox_data.xmax * (new_width / old_width)),
new_width - 1))
bbox_data.ymax = max(
0,
min(int(bbox_data.ymax * (new_height / old_height)),
new_height - 1))
bbox_data.keypoints[:, 0] = (bbox_data.keypoints[:, 0] *
(new_width / old_width)).astype(int)
bbox_data.keypoints[:, 1] = (bbox_data.keypoints[:, 1] *
(new_height / old_height)).astype(int)
bbox_data.keypoints = bbox_data.keypoints.astype(int)
bbox_data.cropped_image = None
keypoints = []
for (x, y) in bbox_data.keypoints:
x = max(0, min(x, new_width - 1))
y = max(0, min(y, new_height - 1))
keypoints.append([x, y])
bbox_data.keypoints = np.array(keypoints).reshape(-1, 2)
for additional_bbox_data in bbox_data.additional_bboxes_data:
resize_coords(additional_bbox_data)
for bbox_data in image_data.bboxes_data:
resize_coords(bbox_data)
image_data.keypoints[:, 0] = (image_data.keypoints[:, 0] *
(new_width / old_width)).astype(int)
image_data.keypoints[:, 1] = (image_data.keypoints[:, 1] *
(new_height / old_height)).astype(int)
keypoints = []
for (x, y) in image_data.keypoints:
x = max(0, min(x, new_width - 1))
y = max(0, min(y, new_height - 1))
keypoints.append([x, y])
image_data.keypoints = np.array(keypoints).reshape(-1, 2)
image_data.image_path = None
image_data.image = image
return image_data
def tf_record_from_image_data(image_data: ImageData,
label_map: Dict[str, int],
use_thumbnail: Tuple[int, int] = None):
filename = image_data.image_path
encoded_filename = str(filename).encode('utf8')
true_bboxes = np.array(
[[bbox_data.xmin, bbox_data.ymin, bbox_data.xmax, bbox_data.ymax]
for bbox_data in image_data.bboxes_data],
dtype=float)
image = image_data.open_image()
height, width, _ = image.shape
if len(true_bboxes) > 0:
normalized_true_bboxes = true_bboxes.copy()
normalized_true_bboxes[:, [0, 2]] /= width
normalized_true_bboxes[:, [1, 3]] /= height
xmins = normalized_true_bboxes[:, 0]
ymins = normalized_true_bboxes[:, 1]
xmaxs = normalized_true_bboxes[:, 2]
ymaxs = normalized_true_bboxes[:, 3]
else:
ymins, xmins, ymaxs, xmaxs = [], [], [], []
encoded_jpg = BytesIO()
image = Image.fromarray(image)
if use_thumbnail:
image.thumbnail(use_thumbnail)
image.save(encoded_jpg, format='JPEG')
encoded_jpg = encoded_jpg.getvalue()
image_format = b'jpg'
class_names = [bbox_data.label for bbox_data in image_data.bboxes_data]
encoded_class_names = [
class_name.encode('utf-8') for class_name in class_names
]
classes = [label_map[class_name] for class_name in class_names]
tf_record = create_tf_record(height=height,
width=width,
encoded_filename=encoded_filename,
encoded_jpg=encoded_jpg,
image_format=image_format,
xmins=xmins,
ymins=ymins,
xmaxs=xmaxs,
ymaxs=ymaxs,
encoded_class_names=encoded_class_names,
classes=classes)
return tf_record
def concat_images_data(image_data_a: ImageData,
image_data_b: ImageData,
background_color_a: Tuple[int, int, int, int] = None,
background_color_b: Tuple[int, int, int, int] = None,
thumbnail_size_a: Tuple[int, int] = None,
thumbnail_size_b: Tuple[int, int] = None,
how: Literal['horizontally',
'vertically'] = 'horizontally',
mode: Literal['L', 'RGB', 'RGBA'] = 'RGBA',
background_edge_width: int = 3,
between_edge_width: int = 0) -> ImageData:
image_data_a = copy.deepcopy(image_data_a)
image_data_b = copy.deepcopy(image_data_b)
if image_data_a is None and image_data_b is not None:
return image_data_b
if image_data_a is not None and image_data_b is None:
return image_data_a
image_a = image_data_a.open_image()
image_b = image_data_b.open_image()
ha, wa = image_a.shape[:2]
hb, wb = image_b.shape[:2]
image = concat_images(image_a=image_a,
image_b=image_b,
background_color_a=background_color_a,
background_color_b=background_color_b,
thumbnail_size_a=thumbnail_size_a,
thumbnail_size_b=thumbnail_size_b,
how=how,
mode=mode,
background_edge_width=background_edge_width,
between_edge_width=between_edge_width)
image_data_a_new_xmin, image_data_a_new_ymin = None, None
image_data_b_new_xmin, image_data_b_new_ymin = None, None
if how == 'horizontally':
max_height = np.max([ha, hb])
min_ha = max_height // 2 - ha // 2
max_ha = max_height // 2 + ha // 2
min_hb = max_height // 2 - hb // 2
max_hb = max_height // 2 + hb // 2
image_data_a_new_xmin = 0
image_data_a_new_ymin = min_ha
image_data_a_new_xmax = wa
image_data_a_new_ymax = max_ha
image_data_b_new_xmin = wa + between_edge_width
image_data_b_new_ymin = min_hb
image_data_b_new_xmax = wa + between_edge_width + wb
image_data_b_new_ymax = max_hb
elif how == 'vertically':
max_width = np.max([wa, wb])
min_wa = max_width // 2 - wa // 2
max_wa = max_width // 2 + wa // 2
min_wb = max_width // 2 - wb // 2
max_wb = max_width // 2 + wb // 2
image_data_a_new_xmin = min_wa
image_data_a_new_ymin = 0
image_data_a_new_xmax = max_wa
image_data_a_new_ymax = ha
image_data_b_new_xmin = min_wb
image_data_b_new_ymin = ha + between_edge_width
image_data_b_new_xmax = max_wb
image_data_b_new_ymax = ha + between_edge_width + hb
keypoints_a = image_data_a.keypoints
keypoints_b = image_data_a.keypoints
keypoints_a[:, 0] += image_data_a_new_xmin
keypoints_a[:, 1] += image_data_a_new_ymin
keypoints_b[:, 0] += image_data_b_new_xmin
keypoints_b[:, 1] += image_data_b_new_ymin
def _get_new_coords_for_bbox_data(bbox_data: BboxData, xmin: int,
ymin: int):
bbox_data.keypoints[:, 0] += xmin
bbox_data.keypoints[:, 1] += ymin
bbox_data.xmin += xmin
bbox_data.ymin += ymin
bbox_data.xmax += xmin
bbox_data.ymax += ymin
bbox_data.image = None
bbox_data.image_path = None
bbox_data.cropped_image = None
for additional_bbox_data in bbox_data.additional_bboxes_data:
_get_new_coords_for_bbox_data(additional_bbox_data, xmin, ymin)
for bbox_data in image_data_a.bboxes_data:
_get_new_coords_for_bbox_data(bbox_data, image_data_a_new_xmin,
image_data_a_new_ymin)
if 'concat_images_data__image_data' not in [
bbox_data.label for bbox_data in image_data_a.bboxes_data
]:
bbox_data_a_into = [
BboxData(xmin=image_data_a_new_xmin,
ymin=image_data_a_new_ymin,
xmax=image_data_a_new_xmax,
ymax=image_data_a_new_ymax,
label='concat_images_data__image_data',
additional_bboxes_data=[
bbox_data for bbox_data in image_data_a.bboxes_data
if 'concat_images_data__image_data' != bbox_data.label
])
]
else:
bbox_data_a_into = []
image_data_a.bboxes_data = [
bbox_data for bbox_data in image_data_a.bboxes_data
if 'concat_images_data__image_data' == bbox_data.label
] + bbox_data_a_into
for bbox_data in image_data_b.bboxes_data:
_get_new_coords_for_bbox_data(bbox_data, image_data_b_new_xmin,
image_data_b_new_ymin)
if 'concat_images_data__image_data' not in [
bbox_data.label for bbox_data in image_data_b.bboxes_data
]:
bbox_data_b_into = [
BboxData(xmin=image_data_b_new_xmin,
ymin=image_data_b_new_ymin,
xmax=image_data_b_new_xmax,
ymax=image_data_b_new_ymax,
label='concat_images_data__image_data',
additional_bboxes_data=[
bbox_data for bbox_data in image_data_b.bboxes_data
if 'concat_images_data__image_data' != bbox_data.label
])
]
else:
bbox_data_b_into = []
image_data_b.bboxes_data = [
bbox_data for bbox_data in image_data_b.bboxes_data
if 'concat_images_data__image_data' == bbox_data.label
] + bbox_data_b_into
image_data = ImageData(image_path=None,
image=image,
bboxes_data=image_data_a.bboxes_data +
image_data_b.bboxes_data,
label=None,
keypoints=np.concatenate([keypoints_a, keypoints_b],
axis=0),
additional_info={
**image_data_a.additional_info,
**image_data_b.additional_info
})
return image_data
def crop_image_data(
image_data: ImageData,
xmin: int,
ymin: int,
xmax: int,
ymax: int,
allow_negative_and_large_coords: bool,
remove_bad_coords: bool,
) -> ImageData:
assert 0 <= xmin and 0 <= ymin
assert xmin <= xmax and ymin <= ymax
image_data = copy.deepcopy(image_data)
image = image_data.open_image()
height, width, _ = image.shape
assert xmax <= width and ymax <= height
image = image[ymin:ymax, xmin:xmax]
new_height, new_width, _ = image.shape
def resize_coords(bbox_data: BboxData):
bbox_data.xmin = bbox_data.xmin - xmin
bbox_data.ymin = bbox_data.ymin - ymin
bbox_data.xmax = bbox_data.xmax - xmin
bbox_data.ymax = bbox_data.ymax - ymin
bbox_data.keypoints[:, 0] -= xmin
bbox_data.keypoints[:, 1] -= ymin
bbox_data.cropped_image = None
if not allow_negative_and_large_coords:
bbox_data.xmin = max(0, min(bbox_data.xmin, new_width - 1))
bbox_data.ymin = max(0, min(bbox_data.ymin, new_height - 1))
bbox_data.xmax = max(0, min(bbox_data.xmax, new_width - 1))
bbox_data.ymax = max(0, min(bbox_data.ymax, new_height - 1))
keypoints = []
for (x, y) in bbox_data.keypoints:
x = max(0, min(x, new_width - 1))
y = max(0, min(y, new_height - 1))
keypoints.append([x, y])
bbox_data.keypoints = np.array(keypoints).reshape(-1, 2)
for additional_bbox_data in bbox_data.additional_bboxes_data:
resize_coords(additional_bbox_data)
for bbox_data in image_data.bboxes_data:
resize_coords(bbox_data)
keypoints = []
for (x, y) in image_data.keypoints:
x = max(0, min(x - xmin, new_width - 1))
y = max(0, min(y - ymin, new_height - 1))
keypoints.append([x, y])
image_data.keypoints = np.array(keypoints).reshape(-1, 2)
def if_bbox_data_inside_crop(bbox_data: BboxData):
bbox_data.keypoints = bbox_data.keypoints[(
(bbox_data.keypoints[:, 0] >= 0) & (bbox_data.keypoints[:, 1] >= 0)
& (bbox_data.keypoints[:, 0] < new_height) &
(bbox_data.keypoints[:, 1] < new_width))]
bbox_data.additional_bboxes_data = [
additional_bbox_data
for additional_bbox_data in bbox_data.additional_bboxes_data
if if_bbox_data_inside_crop(additional_bbox_data)
]
return (bbox_data.xmin >= 0 and bbox_data.ymin >= 0
and bbox_data.xmax < new_width and bbox_data.ymax < new_height
and bbox_data.xmin < bbox_data.xmax
and bbox_data.ymin < bbox_data.ymax)
if remove_bad_coords:
image_data.bboxes_data = [
bbox_data for bbox_data in image_data.bboxes_data
if if_bbox_data_inside_crop(bbox_data)
]
image_data.keypoints = image_data.keypoints[(
(image_data.keypoints[:, 0] >= 0) &
(image_data.keypoints[:, 1] >= 0) &
(image_data.keypoints[:, 0] < new_height) &
(image_data.keypoints[:, 1] < new_width))]
image_data.image_path = None
image_data.image = image
return image_data
def visualize_image_data(
image_data: ImageData,
use_labels: bool = False,
score_type: Literal['detection', 'classification'] = None,
filter_by_labels: List[str] = None,
known_labels: List[str] = None,
draw_base_labels_with_given_label_to_base_label_image: Callable[[str], np.ndarray] = None,
keypoints_radius: int = 5,
include_additional_bboxes_data: bool = False,
additional_bboxes_data_depth: Optional[int] = None,
fontsize: int = 24,
thickness: int = 4,
return_as_pil_image: bool = False
) -> Union[np.ndarray, Image.Image]:
image_data = get_image_data_filtered_by_labels(
image_data=image_data,
filter_by_labels=filter_by_labels
)
image = image_data.open_image()
if include_additional_bboxes_data:
bboxes_data = []
def recursive_get_bboxes_data(bbox_data, depth):
if additional_bboxes_data_depth is not None and depth > additional_bboxes_data_depth:
return
bboxes_data.append(bbox_data)
for bbox_data in bbox_data.additional_bboxes_data:
recursive_get_bboxes_data(bbox_data, depth=depth+1)
for bbox_data in image_data.bboxes_data:
recursive_get_bboxes_data(bbox_data, depth=0)
else:
bboxes_data = image_data.bboxes_data
labels = [bbox_data.label for bbox_data in bboxes_data]
if known_labels is None:
known_labels = list(set(labels))
k_keypoints = [bbox_data.keypoints for bbox_data in bboxes_data]
bboxes = np.array([
(bbox_data.ymin, bbox_data.xmin, bbox_data.ymax, bbox_data.xmax)
for bbox_data in bboxes_data
])
angles = np.array([0. for _ in bboxes_data])
if score_type == 'detection':
scores = np.array([bbox_data.detection_score for bbox_data in bboxes_data])
skip_scores = False
elif score_type == 'classification':
scores = np.array([bbox_data.classification_score for bbox_data in bboxes_data])
skip_scores = False
else:
scores = None
skip_scores = True
image = visualize_boxes_and_labels_on_image_array(
image=image,
bboxes=bboxes,
angles=angles,
scores=scores,
k_keypoints=k_keypoints,
labels=labels,
use_normalized_coordinates=False,
skip_scores=skip_scores,
skip_labels=not use_labels,
groundtruth_box_visualization_color='lime',
known_labels=known_labels,
keypoints_radius=keypoints_radius,
fontsize=fontsize,
thickness=thickness
)
if len(image_data.keypoints) > 0:
image_pil = Image.fromarray(image)
draw = ImageDraw.Draw(image_pil)
for idx, (x, y) in enumerate(image_data.keypoints):
draw.pieslice(
[(x-keypoints_radius, y-keypoints_radius), (x+keypoints_radius, y+keypoints_radius)], start=0, end=360,
fill=STANDARD_COLORS_RGB[idx % len(STANDARD_COLORS_RGB)]
)
image = np.array(image_pil)
if draw_base_labels_with_given_label_to_base_label_image is not None:
for bbox_data in image_data.bboxes_data:
base_label_image = draw_base_labels_with_given_label_to_base_label_image(bbox_data.label)
draw_label_image(
image=image,
base_label_image=base_label_image,
bbox_data=bbox_data,
inplace=True
)
if return_as_pil_image:
return Image.fromarray(image)
return image