def test_denormalize_bboxes():
bboxes = [[0.0375, 0.125, 0.25, 1.0], [0.0375, 0.125, 0.25, 1.0, 99]]
denormalized_bboxes_1 = denormalize_bboxes(bboxes, 200, 400)
denormalized_bboxes_2 = [
denormalize_bbox(bboxes[0], 200, 400),
denormalize_bbox(bboxes[1], 200, 400)
]
assert denormalized_bboxes_1 == denormalized_bboxes_2
def crop_bbox_by_coords(bbox, crop_coords, crop_height, crop_width, rows, cols):
"""Crop a bounding box using the provided coordinates of bottom-left and top-right corners in pixels and the
required height and width of the crop.
"""
bbox = denormalize_bbox(bbox, rows, cols)
x_min, y_min, x_max, y_max = bbox
x1, y1, x2, y2 = crop_coords
cropped_bbox = [x_min - x1, y_min - y1, x_max - x1, y_max - y1]
return normalize_bbox(cropped_bbox, crop_height, crop_width)
def apply_to_bbox(self,
bbox,
pad_top=0,
pad_bottom=0,
pad_left=0,
pad_right=0,
rows=0,
cols=0,
**params):
x_min, y_min, x_max, y_max = denormalize_bbox(bbox, rows, cols)
bbox = x_min + pad_left, y_min + pad_top, x_max + pad_left, y_max + pad_top
return normalize_bbox(bbox, rows + pad_top + pad_bottom,
cols + pad_left + pad_right)
def _convert_to_coco(self, bbox, rows, cols):
if self.source_format == 'pascal_voc':
(x1, y1, x2, y2), tail = bbox[:4], bbox[4:]
return (x1, y1, x2 - x1, y2 - y1) + tail
elif self.source_format == 'yolo':
_bbox, tail = np.array(bbox[:4]), bbox[4:]
if np.any((_bbox <= 0) | (_bbox > 1)):
raise ValueError("In YOLO format all labels must be float "
"and in range (0, 1]")
x, y, width, height = np.round(denormalize_bbox(bbox, rows, cols))
x_min = x - width / 2 + 1
y_min = y - height / 2 + 1
return (x_min, y_min, width, height) + tail
return bbox
def _convert_to_pascal_voc(self, bbox, rows, cols):
(x, y, width, height), tail = bbox[:4], bbox[4:]
if self.source_format == 'coco':
return (x, y, x + width, y + height) + tail
elif self.source_format == 'yolo':
_bbox = np.array(bbox[:4])
if np.any((_bbox <= 0) | (_bbox > 1)):
raise ValueError("In YOLO format all labels must be float "
"and in range (0, 1]")
x, y, width, height = np.round(denormalize_bbox(bbox, rows, cols))
x_min = x - width / 2 + 1
x_max = x_min + width
y_min = y - height / 2 + 1
y_max = y_min + height
return (x_min, y_min, x_max, y_max) + tail
return bbox
def apply_to_bbox(self, bbox, **params):
"""
Removes the bounding boxes which are covered by the applied cutout
:param bbox: A single bounding box coordinates in pascal_voc format
:returns transformed bbox's coordinates
"""
# Denormalize the bbox coordinates
bbox = denormalize_bbox(bbox, self.img_height, self.img_width)
x_min, y_min, x_max, y_max = tuple(map(int, bbox))
bbox_size = (x_max - x_min) * (y_max - y_min) # width * height
overlapping_size = np.sum(
(self.image[y_min:y_max, x_min:x_max, 0] == self.fill_value)
& (self.image[y_min:y_max, x_min:x_max, 1] == self.fill_value)
& (self.image[y_min:y_max, x_min:x_max, 2] == self.fill_value))
# Remove the bbox if it has more than some threshold of content is inside the cutout patch
if overlapping_size / bbox_size > self.bbox_removal_threshold:
return normalize_bbox((0, 0, 0, 0), self.img_height,
self.img_width)
return normalize_bbox(bbox, self.img_height, self.img_width)
def test_denormalize_normalize_bbox(bbox):
denormalized_bbox = denormalize_bbox(bbox, 200, 400)
normalized_bbox = normalize_bbox(denormalized_bbox, 200, 400)
assert normalized_bbox == bbox
def test_normalize_denormalize_bbox(bbox):
normalized_bbox = normalize_bbox(bbox, 200, 400)
denormalized_bbox = denormalize_bbox(normalized_bbox, 200, 400)
assert denormalized_bbox == bbox
def test_denormalize_bbox(bbox, expected):
denormalized_bbox = denormalize_bbox(bbox, 200, 400)
assert denormalized_bbox == expected
