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 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 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 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