def _get_annotation(item, classes):
# Annotation is encoded in the original image size, actual image size may be smaller due to performance reasons
width, height = item['image_size']
width, height = float(width), float(height)
annotation = {}
for obj in item['objects']:
class_name = obj['label']
if class_name == 'ignored':
continue
try:
class_id = classes.index(class_name)
except ValueError:
tf.logging.warning(
"Unknown label = '{0}', supported labels: {1}".format(
class_name, classes))
class_id = -1
continue
xmin, ymin, xmax, ymax = obj['bbox']
xmin /= width
ymin /= height
xmax /= width
ymax /= height
annotation.setdefault(class_id, []).append(
BoundingBox(xmin, ymin, xmax, ymax))
return annotation
def _transform_image(self, image):
"""
Geometrically transforms image according to parameters
:type image: input image
:rtype: transformed image
"""
top_h, top_w = self._infer_top_shape(image)
scale = self.transform_param.scale
crop_w, crop_h = self.transform_param.crop_size
do_mirror = self.transform_param.mirror and random.randint(0, 1)
if self.transform_param.resize_param:
image = apply_resize(image, self.transform_param.resize_param)
if self.transform_param.noise_param:
image = apply_noise(image, self.transform_param.noise_param)
image_h, image_w = image.shape[0:2]
w_off, h_off = (0, 0)
if crop_h * crop_w > 0:
if self.is_training:
def exclusive_random(val):
return random.randint(0, val - 1)
h_off = exclusive_random(image_h - crop_h + 1)
w_off = exclusive_random(image_w - crop_w + 1)
else:
h_off = (image_h - crop_h) / 2
w_off = (image_w - crop_w) / 2
image = image[h_off:h_off + crop_h, w_off:w_off + crop_w]
xmin = float(w_off) / image_w
ymin = float(h_off) / image_h
xmax = float(w_off + top_w) / image_w
ymax = float(h_off + top_h) / image_h
crop_bbox = BoundingBox(xmin, ymin, xmax, ymax)
if do_mirror:
image = cv2.flip(image, 1)
if self.transform_param.mean_value:
image = image.astype(np.float32)
image -= self.transform_param.mean_value
if math.fabs(scale - 1) > 1e-2:
image = image.astype(np.float32)
image *= scale
return image, crop_bbox, do_mirror
def update_bbox_by_resize_policy(old_width, old_height, bbox, resize_param):
new_height = resize_param.height
new_width = resize_param.width
orig_aspect = float(old_width) / old_height
new_aspect = new_width / new_height
x_min = bbox.xmin * old_width
y_min = bbox.ymin * old_height
x_max = bbox.xmax * old_width
y_max = bbox.ymax * old_height
if resize_param.resize_mode == ResizeParameter.WARP:
x_min = max(0., x_min * new_width / old_width)
x_max = min(new_width, x_max * new_width / old_width)
y_min = max(0., y_min * new_height / old_height)
y_max = min(new_height, y_max * new_height / old_height)
elif resize_param.resize_mode == ResizeParameter.FIT_LARGE_SIZE_AND_PAD:
if orig_aspect > new_aspect:
padding = (new_height - new_width / orig_aspect) / 2
x_min = max(0., x_min * new_width / old_width)
x_max = min(new_width, x_max * new_width / old_width)
y_min = y_min * (new_height - 2 * padding) / old_height
y_min = padding + max(0., y_min)
y_max = y_max * (new_height - 2 * padding) / old_height
y_max = padding + min(new_height, y_max)
else:
padding = (new_width - orig_aspect * new_height) / 2
x_min = x_min * (new_width - 2 * padding) / old_width
x_min = padding + max(0., x_min)
x_max = x_max * (new_width - 2 * padding) / old_width
x_max = padding + min(new_width, x_max)
y_min = max(0., y_min * new_height / old_height)
y_max = min(new_height, y_max * new_height / old_height)
elif resize_param.resize_mode == ResizeParameter.FIT_SMALL_SIZE:
if orig_aspect < new_aspect:
new_height = new_width / orig_aspect
else:
new_width = orig_aspect * new_height
x_min = max(0., x_min * new_width / old_width)
x_max = min(new_width, x_max * new_width / old_width)
y_min = max(0., y_min * new_height / old_height)
y_max = min(new_height, y_max * new_height / old_height)
result = BoundingBox(difficult=bbox.difficult)
result.xmin = x_min / new_width
result.ymin = y_min / new_height
result.xmax = x_max / new_width
result.ymax = y_max / new_height
return result
def expand_image(image, expand_ratio, mean_value=None):
height, width, chs = image.shape
# Get the bbox dimension.
new_h = int(height * expand_ratio)
new_w = int(width * expand_ratio)
h_off = int(math.floor(random.uniform(0., new_h - height)))
w_off = int(math.floor(random.uniform(0., new_w - width)))
xmin = -w_off / float(width)
ymin = -h_off / float(height)
xmax = (new_w - w_off) / float(width)
ymax = (new_h - h_off) / float(height)
expand_box = BoundingBox(xmin, ymin, xmax, ymax)
fill_value = mean_value if mean_value is not None else [0] * chs
expanded_image = np.full(shape=(new_h, new_w, chs),
fill_value=fill_value,
dtype=image.dtype)
expanded_image[h_off:h_off + height, w_off:w_off + width, :] = image
return expanded_image, expand_box