def generate_training_list():
file_list = glob_files(args.db_dir, ['.png'])
file_list = [f for f in file_list if 'texture' in os.path.basename(f)]
[os.makedirs(os.path.dirname(f.replace(args.db_dir, args.out_dir))) for f in file_list if not
os.path.exists(os.path.dirname(f.replace(args.db_dir, args.out_dir)))]
with Pool() as p:
content = list(tqdm(p.imap_unordered(crop_image, file_list), total=len(file_list)))
def detect_face():
files = glob_files(args.db_dir, ['.jpg', '.jpeg', '.png', '.bmp'])
detector = dlib.cnn_face_detection_model_v1(args.dlib_model)
contents = []
p_bar = tqdm.tqdm(total=len(files))
for f in files:
ext = os.path.splitext(f)[1]
file_bbox = f.replace(args.db_dir, args.out_dir)
file_bbox = file_bbox.replace(ext, '.roi')
file_landmark = f.replace(args.db_dir, args.lm_dir)
file_landmark = file_landmark.replace(ext, args.lm_ext)
if os.path.exists(file_bbox):
contents.append([f, file_bbox])
else:
if not os.path.exists(os.path.dirname(file_bbox)):
os.makedirs(os.path.dirname(file_bbox))
im = cv2.imread(f)
if args.use_lm:
box = gen_bbox_from_landmarks(file_landmark)
else:
box = detect_dlib(detector, f)
if len(box) == 0:
# center crop
box = center_crop(im)
if args.plot:
plt.figure()
plt.imshow(cv2.cvtColor(im, cv2.COLOR_BGR2RGB))
ax = plt.gca()
ax.add_patch(patches.Rectangle((box[0], box[1]), box[2], box[3], fill=False, ec='g'))
plt.show()
with open(file_bbox, 'w') as fout:
fout.write('%d %d %d %d' % (box[0], box[1], box[2], box[3]))
contents.append([f, file_bbox])
p_bar.update()
p_bar.close()
if len(args.out_path) > 0:
with open(args.out_path, 'w') as f:
csv_writer = csv.writer(f)
csv_writer.writerow(['FILE', 'ROI'])
for im_path, roi_path in contents:
csv_writer.writerow([im_path.replace(args.db_dir, ''), os.path.realpath(roi_path)])
def merge_im_box():
im_files = glob_files(args.db_dir, ['.jpg', '.jpeg', '.png', '.bmp'])
box_files = [
im_file.replace(args.db_dir, args.box_dir).replace(
os.path.splitext(im_file)[1], args.box_ext) for im_file in im_files
]
content = [[im_file.replace(args.db_dir, ''), box_file]
for im_file, box_file in zip(im_files, box_files)]
with open(args.out_csv, 'w') as f:
csv_writer = csv.writer(f)
csv_writer.writerow(['FILE', 'ROI'])
for line in content:
csv_writer.writerow(line)
def write_final_csv():
files = glob_files(args.out_dir, ['.png'])
p_bar = tqdm(total=len(files))
with open(os.path.join(args.out_dir, 'IJB-%s.csv' % args.set), 'w') as f:
csv_writer = csv.writer(f)
csv_writer.writerow(['FILE', 'ROI'])
for file_path in files:
box_path = os.path.realpath(file_path).replace('.png', '.roi')
csv_writer.writerow(
[file_path.replace(args.out_dir, ''), box_path])
p_bar.update()
p_bar.close()
def extracting_features():
# glob images
im_paths = glob_files(args.db_dir, args.im_exts)
feature_paths = [
im_path.replace(args.db_dir, args.ft_dir).replace(
os.path.splitext(os.path.basename(im_path))[1], args.ft_ext)
for im_path in im_paths
]
# generate feature directory
[
os.makedirs(os.path.dirname(feature_path))
for feature_path in feature_paths
if not os.path.exists(os.path.dirname(feature_path))
]
# loading model
use_symbol_block = False
if len(args.symbol_prefix) > 0:
use_symbol_block = True
sym, arg_params, aux_params = mx.model.load_checkpoint(
args.symbol_prefix, 0)
internals = sym.get_internals()
inputs = internals['data']
outputs_blobs = [
internals[layer_name + '_output']
for layer_name in args.output_blobs.split(',')
]
inference = gluon.SymbolBlock(outputs_blobs, inputs)
else:
inference = gluon.model_zoo.vision.get_model(args.arch,
classes=args.num_classes)
inference.load_params(args.weights, ctx=ctx)
if len(args.symbol_prefix) == 0:
inference.hybridize()
# extracting features
global_start_time = timeit.default_timer()
valid_counts = 0
for im_idx, (im_path,
feature_path) in enumerate(zip(im_paths, feature_paths)):
if not os.path.exists(feature_path):
start_time = timeit.default_timer()
im = image_processing(im_path)
if im is not None:
im = im.as_in_context(ctx)
elapsed_time = timeit.default_timer()
processing_time = elapsed_time - start_time
start_time = timeit.default_timer()
if use_symbol_block:
features = inference(im)
else:
features = inference.features(im)
feature_extracting_time = timeit.default_timer() - start_time
features = features.asnumpy().flatten()
feature = Signature(features)
feature.save_features(feature_path)
valid_counts += 1
print(
'Presssed [%d/%d]: %s \t Pre-processing time: %.2f ms\t Extracting features time: %.2f ms'
% (im_idx, len(im_paths), im_path, processing_time * 1000,
feature_extracting_time * 1000))
else:
print('The feature file exists, skip the file')
global_elapsed_time = timeit.default_timer() - global_start_time
print(
'Total elapsed time: %s \t Processed [%d] images \t Avg. time: %.2f ms'
% (str(datetime.timedelta(seconds=global_elapsed_time)), valid_counts,
global_elapsed_time * 1000 / valid_counts))
def main():
if args.set == 'A':
files = glob_files(os.path.join(args.db_dir, 'IJB-A_1N_sets'),
['.csv'])
elif args.set == 'B':
files = [os.path.join(args.db_dir, 'protocol/ijbb_face_detection.csv')]
elif args.set == 'C':
files = [
os.path.join(args.db_dir,
'protocol/ijbc_face_detection_ground_truth.csv')
]
else:
raise NotImplementedError
contents = []
idx_tmp_id, idx_sub_id, idx_file, idx_x, idx_y, idx_w, idx_h, idx_reye_x, idx_reye_y, idx_leye_x, idx_leye_y, \
idx_nose_x, idx_nose_y = 0, 1, 2, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
for file_ind, file_path in enumerate(files):
print('Processing [%d]: %s...' % (file_ind, file_path))
with open(file_path) as f:
csv_reader = csv.reader(f)
for idx, row in enumerate(csv_reader):
if idx == 0:
idx_tmp_id = row.index('TEMPLATE_ID')
idx_sub_id = row.index('SUBJECT_ID')
idx_file = row.index('FILE')
idx_x = row.index('FACE_X')
idx_y = row.index('FACE_Y')
idx_w = row.index('FACE_WIDTH')
idx_h = row.index('FACE_HEIGHT')
idx_reye_x = row.index('RIGHT_EYE_X')
idx_reye_y = row.index('RIGHT_EYE_Y')
idx_leye_x = row.index('LEFT_EYE_X')
idx_leye_y = row.index('LEFT_EYE_Y')
idx_nose_x = row.index('NOSE_BASE_X')
idx_nose_y = row.index('NOSE_BASE_Y')
else:
template_id = row[idx_tmp_id]
subject_id = row[idx_sub_id]
im_path = row[idx_file]
x, y, w, h = float(row[idx_x]), float(row[idx_y]), float(
row[idx_w]), float(row[idx_h])
reye_x, reye_y, leye_x, leye_y, nose_x, nose_y = row[idx_reye_x], row[idx_reye_y], row[idx_leye_x],\
row[idx_leye_y], row[idx_nose_x], row[idx_nose_y]
if len(reye_x) > 0 and len(reye_y) > 0 and len(
leye_x) > 0 and len(leye_y) > 0 and len(
nose_x) > 0 and len(nose_y) > 0:
reye_x, reye_y, leye_x, leye_y, nose_x, nose_y = float(reye_x), float(reye_y), float(leye_x), \
float(leye_y), float(nose_x), float(nose_y)
c_x, c_y = nose_x, nose_y
radius = math.sqrt(
max((reye_x - nose_x)**2 + (reye_y - nose_y)**2,
(leye_x - nose_x)**2 + (leye_y - nose_y)**2))
radius *= 1.1
x, y = c_x - radius, c_y - radius
w, h = 2 * radius, 2 * radius
im_name = os.path.splitext(os.path.basename(im_path))[0]
dir_name = os.path.dirname(im_path)
out_im_path = os.path.join(args.out_dir, dir_name,
'%s.png' % im_name)
if not os.path.exists(os.path.dirname(out_im_path)):
os.makedirs(os.path.dirname(out_im_path))
if len(args.box_dir) > 0:
out_box_path = os.path.join(args.box_dir, dir_name,
'%s.roi' % im_name)
if not os.path.exists(os.path.dirname(out_box_path)):
os.makedirs(os.path.dirname(out_box_path))
else:
out_box_path = os.path.join(args.out_dir, dir_name,
'%s.roi' % im_name)
contents.append([
im_path, template_id, subject_id,
int(x),
int(y),
int(w),
int(h), out_im_path, out_box_path
])
with Pool() as p:
processed_contents = list(
tqdm(p.imap(crop_image, contents), total=len(contents)))
processed_contents = list(filter(None, processed_contents))
processed_contents = remove_duplicates(processed_contents)
if len(args.out_path) > 0:
if not os.path.exists(os.path.dirname(args.out_path)):
os.makedirs(os.path.dirname(args.out_path))
write_csv(args.out_path, processed_contents)