def unary_feature_extraction(anno_dir, photo_dir, output_dir, label_dict_file=None):
"""
features extraction to train unary potential
"""
# prepare gabor kernels
gabor_kernels = feat.prepare_gabor_kernels(gabor_freqs=[0.2])
# prepare face detectors
face_detectors = feat.prepare_face_detectors()
# label dictionary
label_dict = util.get_label_dictionary(label_dict_file)
# get all csv annotation files
csv_dirs = os.listdir(anno_dir)
X, y = np.array([]), np.array([])
for i, csv_file in enumerate(csv_dirs):
print "%i from %i" % (i, len(csv_dirs))
csv_path = os.path.join(anno_dir, csv_file)
image_path = os.path.join(photo_dir, csv_file.split(".")[0] + ".jpg")
if os.path.isfile(csv_path):
image = io.imread(image_path)
sps = feat.compute_superpixels(image, 300)
pixel_class = feat.load_pixel_annotations(csv_path)
unary_features = feat.compute_unary_features(image, sps, gabor_kernels, face_detectors, pixel_class)
if unary_features is not None:
# generate labels
labels = feat.generate_labels(pixel_class, sps, label_dict)
X = np.vstack([X, unary_features]) if X.size else unary_features
y = np.append(y, labels) if y.size else labels
# save the features
part_idx = anno_dir.split("/")[-2]
np.save(os.path.join(output_dir, 'X_unary_part_' + part_idx + '.npy'), X)
np.save(os.path.join(output_dir, 'y_unary_part_' + part_idx + '.npy'), y)
def crf_feature_extraction(anno_dir,
photo_dir,
unary_scaler,
unary_clf,
output_dir,
label_dict_file=None):
"""
features extraction to train unary potential
"""
# prepare gabor kernels
gabor_kernels = feat.prepare_gabor_kernels(gabor_freqs=[0.2])
# prepare face detectors
face_detectors = feat.prepare_face_detectors()
# label dictionary
label_dict = util.get_label_dictionary(label_dict_file)
X, y = [], []
csv_dirs = os.listdir(anno_dir)
for i, csv_file in enumerate(csv_dirs):
print "%i from %i" % (i, len(csv_dirs))
csv_path = os.path.join(anno_dir, csv_file)
image_path = os.path.join(photo_dir, csv_file.split(".")[0] + ".jpg")
if os.path.isfile(csv_path):
image = io.imread(image_path)
sps = feat.compute_superpixels(image, 300)
pixel_class = feat.load_pixel_annotations(csv_path)
unary_features = feat.compute_unary_features(
image, sps, gabor_kernels, face_detectors, pixel_class)
if unary_features is not None:
# compute crf node features from unary potentials
scaled_unary_features = unary_scaler.transform(unary_features)
node_features = unary_clf.predict_proba(scaled_unary_features)
# generate edges
# vertices, edges = feat.generate_edges(sps)
edges = np.array(graph.rag_mean_color(image, sps).edges())
# extract edge features
edge_features = feat.compute_crf_edge_features(
unary_features, edges)
# generate labels
labels = feat.generate_labels(pixel_class, sps, label_dict)
X.append((node_features, edges, edge_features))
y.append(labels)
# save the features
part_idx = anno_dir.split("/")[-2]
np.save(os.path.join(output_dir, 'X_crf_part_' + part_idx + '.npy'), X)
np.save(os.path.join(output_dir, 'y_crf_part_' + part_idx + '.npy'), y)
DEFAULT_MAX_WIDTH = 100
if __name__ == "__main__":
# argument parser
parser = argparse.ArgumentParser()
parser.add_argument('--features', help='features filename')
parser.add_argument('--image-query', help='image query path')
args = parser.parse_args()
# load features
(tree, image_paths) = pickle.load(open(args.features, 'rb'))
# prepare gabor kernels
gabor_kernels = feat.prepare_gabor_kernels(gabor_freqs=[0.2])
# load image
extension = "." + args.image_query.split(".")[-1]
crop_path = args.image_query.split(extension)[0] + "_crop" + extension
image = io.imread(crop_path)
# compute scale factor
h, w, d = image.shape
scale = float(DEFAULT_MAX_WIDTH) / w
# resize image
image = tran.rescale(image, scale=(scale, scale))
image_gray = color.rgb2gray(image)
# load mask