def main():
dataset = args.dataset
root_dir = args.root_dir
filenamelist = get_list_from_filenames(args.filename_list)
face_detector = FaceDetector(detection_cfg)
count = 0
det_count = 0
faceboxes = {}
if dataset == 'BIWI':
img_ext = '_rgb.png'
else:
img_ext = '.jpg'
for image_name in filenamelist:
image_path = os.path.join(root_dir, image_name + img_ext)
image = cv2.imread(image_path)
bboxes = face_detector.detect_faces(image)
count += 1
if len(bboxes) > 0:
det_count += 1
print("images, %s, detected face %s" % (count, det_count))
x_min, y_min, x_max, y_max = bboxes[0][:4]
xmin = int(x_min)
ymin = int(y_min)
xmax = int(x_max)
ymax = int(y_max)
bbox = [xmin, ymin, xmax, ymax]
faceboxes[image_name] = bbox
# if count >10:
# break
headpose_file = os.path.join('datasets', 'filenamelists',
dataset + '_facebbox_disgard.json')
with open(headpose_file, 'w') as f:
json.dump(faceboxes, f, indent=4)
return
def main():
"""Main entrypoint when running this module from the terminal."""
parser = argparse.ArgumentParser(
description='Process and a clean a raw dataset of halloween costumes')
parser.add_argument(
'dataset_source',
help='The path to the directory containing the source dataset.',
type=Path,
action=ReadableDirectory)
parser.add_argument(
'--destination',
'-d',
dest='dataset_destination',
help='The path to the directory which the cleaned '
'dataset should be saved. If this is not specified, the cleaned files are saved in the same parent '
'folder as the source.',
type=Path,
default=None)
parser.add_argument(
'--file-glob-patterns',
nargs='+',
type=str,
default=['*.png', '*.jpeg', '*.jpg'],
help='The glob patterns to use to find files in the source directory.')
parser.add_argument(
'--no-remove-transparency',
action='store_false',
dest='remove_transparency',
help='Remove transparency and replace it with a colour.')
parser.add_argument('--bg-colour',
type=str,
default='WHITE',
help='The colour to replace transparency with.')
parser.add_argument(
'--u2net-size',
type=str,
default='large',
help=
'The size of the pretrained U-2-net model. Either \'large\' or \'small\'.'
)
parser.add_argument(
'--face_image_ratio_threshold',
type=float,
default=0.05,
help='The maximum face-to-image area ratio that is allowed.')
parser.add_argument('--crop-faces',
dest='crop_faces',
action='store_true',
help='Crop out faces.')
parser.add_argument('--yes', '-y', action='store_true', help='Yes to all.')
args = parser.parse_args()
# Create a destination path if none was provided.
if args.dataset_destination is None:
args.dataset_destination = args.dataset_source.parent / (
args.dataset_source.stem + '_cleaned')
if args.dataset_destination.exists() and any(
args.dataset_destination.iterdir()):
if not args.yes:
click.confirm(
f'The destination path (\'{args.dataset_destination.resolve()}\') '
'already exists! Would you like to continue? This will overwrite the directory.',
abort=True)
_rmtree(args.dataset_destination)
args.dataset_destination.mkdir(exist_ok=True, parents=True)
u2net = U2Net(pretrained_model_name=args.u2net_size)
face_detector = FaceDetector()
files = list(get_files(args.dataset_source, args.file_glob_patterns))
with tqdm.tqdm(files) as progress:
for file in progress:
progress.set_description(f'Processing {file.name}')
# Skip images that don't have a single face in them...
face_detection_results = face_detector.detect_faces(file)
if len(face_detection_results) != 1:
continue
segmentation_map = u2net.segment_image(file)
try:
# Remove background from image (using U2Net)
image = u2net.remove_background(file, segmentation_map)
old_image_width, old_image_height = image.size
except InvalidImageError as e:
continue
# Crop image to bounding box (using U2Net)
bounding_box = u2net.get_bounding_box(segmentation_map)
image = image.crop(bounding_box)
# Tuple of the form (x1, y1, x2, y2)
fbb = face_detection_results[0].bounding_box
fbb_width = (fbb[2] - fbb[0])
fbb_height = (fbb[3] - fbb[1])
image_width, image_height = image.size
# Compute the face-to-image area ratio.
# This is used as a heuristic to filter out portrait images
# (i.e. when the face takes up more than a certain percentage of the total image).
face_image_ratio = (fbb_width * fbb_height) / (image_width *
image_height)
if face_image_ratio > args.face_image_ratio_threshold:
continue
if args.crop_faces:
# Crop out the face...
# This assumes that the image is of a person standing vertically.
# Convert the bottom-right y-coordinate of the face bounding box
# into the coordinate system AFTER cropping.
adjusted_fbb_y2 = fbb[3] - bounding_box[1]
image = image.crop((0, adjusted_fbb_y2 - fbb_height * 0.10,
image_width, image_height))
if args.remove_transparency:
# Replace transparency with colour
background_image = Image.new('RGBA', image.size,
args.bg_colour)
background_image.paste(image, (0, 0), image)
image = background_image.convert('RGB')
# Output processed image
destination = args.dataset_destination / (file.stem + '.png')
image.save(str(destination))