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 non_max_suppression_image_data(image_data: ImageData,
iou: float) -> ImageData:
image_data = copy.deepcopy(image_data)
current_bboxes_data = image_data.bboxes_data.copy()
new_bboxes_data = []
while len(current_bboxes_data) != 0:
current_bbox_data = current_bboxes_data[0]
success = True
if len(current_bboxes_data) > 1:
for idx, bbox_data in enumerate(current_bboxes_data):
if idx == 0:
continue
bbox_iou = intersection_over_union(current_bbox_data,
bbox_data)
if bbox_iou >= iou:
pairs_bboxes_data = [bbox_data, current_bbox_data]
pairs_scores = [
possible_bbox_data.detection_score if
possible_bbox_data.detection_score is not None else 1.
for possible_bbox_data in pairs_bboxes_data
]
top_score_idx = np.argmax(pairs_scores)
current_bboxes_data.pop(idx)
current_bboxes_data.pop(0)
current_bboxes_data.append(
BboxData(
xmin=min(bbox_data.xmin, current_bbox_data.xmin),
ymin=min(bbox_data.ymin, current_bbox_data.ymin),
xmax=max(bbox_data.xmax, current_bbox_data.xmax),
ymax=max(bbox_data.ymax, current_bbox_data.ymax),
detection_score=pairs_bboxes_data[top_score_idx].
detection_score,
label=pairs_bboxes_data[top_score_idx].label,
keypoints=pairs_bboxes_data[top_score_idx].
keypoints,
additional_bboxes_data=pairs_bboxes_data[
top_score_idx].additional_bboxes_data,
additional_info=pairs_bboxes_data[top_score_idx].
additional_info))
success = False
break
if success:
new_bboxes_data.append(current_bboxes_data.pop(0))
image_data.bboxes_data = new_bboxes_data
image_data.image_path = image_data.image_path
image_data.image = image_data.image
return image_data
def flatten_additional_bboxes_data_in_image_data(
image_data: ImageData,
additional_bboxes_data_depth: Optional[int] = None,
) -> ImageData:
image_data = copy.deepcopy(image_data)
bboxes_data = []
def _append_bbox_data(bbox_data: BboxData, depth: int):
if additional_bboxes_data_depth is not None and depth > additional_bboxes_data_depth:
return
bboxes_data.append(bbox_data)
for additional_bbox_data in bbox_data.additional_bboxes_data:
_append_bbox_data(additional_bbox_data)
bbox_data.additional_bboxes_data = []
for bbox_data in bboxes_data:
_append_bbox_data(bbox_data, depth=0)
image_data.bboxes_data = bboxes_data
return image_data
def non_max_suppression_image_data_using_tf(image_data: ImageData,
iou: float) -> ImageData:
import tensorflow as tf
image_data = copy.deepcopy(image_data)
if len(image_data.bboxes_data) <= 1:
return image_data
bboxes = [(bbox_data.ymin, bbox_data.xmin, bbox_data.ymax, bbox_data.xmax)
for bbox_data in image_data.bboxes_data]
scores = [
bbox_data.detection_score
if bbox_data.detection_score is not None else 1.
for bbox_data in image_data.bboxes_data
]
result = tf.image.non_max_suppression(bboxes,
scores,
len(image_data.bboxes_data),
iou_threshold=iou)
image_data.bboxes_data = [
image_data.bboxes_data[i] for i in result.numpy()
]
return image_data
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