def main():
args = parse_args()
## MOTION FILTER + SKELETON NORM (SAVE NUMPY)
videos_dirs = get_videos_path(args.dataset_path)
for video_dir in tqdm(videos_dirs, desc='Processing videos...'):
## Create folder for .npy files
os.makedirs(video_dir + '/data/', exist_ok=True)
## Get jsons files
jsons_path = video_dir + '/openpose/json/'
jsons_files = os.listdir(jsons_path)
jsons_files = sorted(jsons_files, key=sort_jsons)
frames = load_video_json(jsons_files, jsons_path)
frames = treat_missing_joints(frames)
frames = np.array( [normalize_kp(frames[frame, :, :]) for frame in range(frames.shape[0])] )
frames = motion_filter(frames)
n_samples, _, _ = frames.shape
for sample in range(n_samples):
np.save(video_dir + '/data/' + jsons_files[sample].split('.json')[0] + '.npy', frames[sample, :, :])
def main():
args = parse_args()
videos_dir = get_videos_path(args.dataset_path)
videos_path = [video_dir + '/' + video_dir.split('/')[-1] + '.mp4' for video_dir in videos_dir]
for idx, video_path in enumerate(tqdm(videos_path, desc="Rendering debug videos...")):
video_name = video_path.split('/')[-1]
filter_video(video_name, videos_dir[idx])
def main():
args = parse_args()
videos_path = get_videos_path(args.dataset_path)
videos_path = [
video_path + '/' + video_path.split('/')[-1] + '.mp4'
for video_path in videos_path
]
for video in tqdm(videos_path, desc="Standardizing videos..."):
standardize_data(video, args.fps, args.sample_rate)
def main():
args = parse_args()
videos_path = get_videos_path(args.dataset_path)
videos_path = [
video_path + '/' + video_path.split('/')[-1] + '.mp4'
for video_path in videos_path
]
for video in tqdm(videos_path, desc="Extracting audio from videos..."):
extract_audio(video)
def main():
args = parse_args()
videos_dirs = get_videos_path(args.dataset_path)
videos_paths = [video_dir + '/' + video_dir.split('/')[-1] + '.mp4' for video_dir in videos_dirs]
txts_paths = [video_dir + '/' + video_dir.split('/')[-1] + '.txt' for video_dir in videos_dirs]
for idx, video_path in enumerate(tqdm(videos_paths, desc="Cutting videos...")):
video_intervals = open(txts_paths[idx], 'r').read().splitlines()
video_intervals = [video_interval.replace(' ', '') for video_interval in video_intervals if video_interval != '']
video_intervals = [( int(interval.split(',')[0]) , int(interval.split(',')[1]) ) for interval in video_intervals]
style_dir = videos_dirs[idx][:videos_dirs[idx].rfind('/')] + '/'
split_and_organize_data(video_intervals, video_path, style_dir)
def main():
args = parse_args()
videos_dir = get_videos_path(args.dataset_path)
videos_path = [
video_dir + '/' + video_dir.split('/')[-1] + '.mp4'
for video_dir in videos_dir
]
for idx, video_path in enumerate(
tqdm(videos_path, desc="Filtering openpose predictions...")):
video_name = video_path.split('/')[-1]
json_files = os.listdir(videos_dir[idx] + '/openpose/json/')
json_files = sorted(json_files, key=sort_openpose_jsons)
filter_video(video_name, json_files, videos_dir[idx])
def main():
args = parse_args()
'''Params arguments
(joints_to_aug, noise_x, noise_y, sigma, scale, flip, noise)'''
params = [ # 40 # 30
([3, 4], True, True, 9.765625, 30, False, True),
([6, 7], True, True, 9.765625, 30, False, True),
([10, 11], True, True, 9.765625, 30, False, True),
([13, 14], True, True, 9.765625, 30, False, True),
([3, 4, 6, 7], True, True, 9.765625, 30, False, True),
([10, 11, 13, 14], True, True, 9.765625, 30, False, True),
([3, 4, 10, 11], True, True, 9.765625, 30, False, True),
([6, 7, 13, 14], True, True, 9.765625, 30, False, True),
([3, 4, 13, 14], True, True, 9.765625, 30, False, True),
([6, 7, 10, 11], True, True, 9.765625, 30, False, True),
([3, 4], True, True, 9.765625, 40, False, True),
([6, 7], True, True, 9.765625, 40, False, True),
([10, 11], True, True, 9.765625, 40, False, True),
([13, 14], True, True, 9.765625, 40, False, True),
([3, 4, 6, 7], True, True, 9.765625, 40, False, True),
([10, 11, 13, 14], True, True, 9.765625, 40, False, True),
([3, 4, 10, 11], True, True, 9.765625, 40, False, True),
([6, 7, 13, 14], True, True, 9.765625, 40, False, True),
([3, 4, 13, 14], True, True, 9.765625, 40, False, True),
([6, 7, 10, 11], True, True, 9.765625, 40, False, True)
]
videos_dir = get_videos_path(args.dataset_path)
videos_dir = [
video_dir for video_dir in videos_dir
if len(video_dir.split('/')[-1].split('_')) == 2
]
if args.pose:
for i, param in enumerate(params):
print("Processing dataset with params config " + str(i + 1) + "/" +
str(len(params)))
augment_data(videos_dir, args, param, i)
if args.audio:
print("Augmenting audio data...")
augment_audio(videos_dir, args, params)
def main():
args = parse_args()
videos_dir = get_videos_path(args.dataset_path)
if args.data_aug:
wavs_paths = [
video_dir + '/' + video_dir.split('/')[-1] + '.wav'
for video_dir in videos_dir
]
#wavs_paths = [re.sub(r'(I_\d+)_\d+', r'\1', wav_path) for wav_path in wavs_paths] # Uncomment this line if you used audio data augmentation
videos_paths = [
video_dir + '/' + video_dir.split('/')[-1] + '.mp4'
for video_dir in videos_dir
]
videos_paths = [
re.sub(r'(I_\d+)_\d+', r'\1', video_path)
for video_path in videos_paths
]
else:
## Remove data augmentation dirs
videos_dir = [
video_dir for video_dir in videos_dir
if len(video_dir.split('/')[-1].split('_')) == 2
]
wavs_paths = [
video_dir + '/' + video_dir.split('/')[-1] + '.wav'
for video_dir in videos_dir
]
videos_paths = [
video_dir + '/' + video_dir.split('/')[-1] + '.mp4'
for video_dir in videos_dir
]
videos_style = [video_dir.split('/')[-2] for video_dir in videos_dir]
samples_dict = {}
global_sample_idx = 0
for idx, video_dir in enumerate(
tqdm(videos_dir, desc='Processing videos...')):
data_path = video_dir + '/data/'
data_files = os.listdir(data_path)
data_files = sorted(data_files, key=sort_npy)
skeletons_dict, indices_dict = get_skeletons(data_files, data_path,
args.stride,
args.sample_size)
## Get audio sample rate
wav_path = wavs_paths[idx]
sample_rate = librosa.get_samplerate(wav_path)
## Get video fps
video_path = videos_paths[idx]
cap = cv2.VideoCapture(video_path)
fps = round(cap.get(cv2.CAP_PROP_FPS))
## Map npy files to audio intervals/sequences
audio_dict = map_video_to_audio(indices_dict, fps, sample_rate)
## Saving info
for sample_idx, skeletons_sample in skeletons_dict.items():
audio_sample = audio_dict[sample_idx]
metadata = {}
metadata['samples_frames'] = skeletons_sample
metadata['samples_audio'] = audio_sample
metadata['audio_path'] = wav_path
metadata['style'] = videos_style[idx]
metadata['video_id'] = idx
samples_dict[global_sample_idx] = metadata
global_sample_idx = global_sample_idx + 1
## Saving pickle file
with open(
args.dataset_path + 'metadata_' + str(args.sample_size) + '_' +
str(args.stride) + '_' + str(args.data_aug) + '.pickle',
'wb') as f:
pickle.dump(samples_dict, f, protocol=pickle.HIGHEST_PROTOCOL)