1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
|
import argparse
import glob
import math
import os
import shutil
import warnings
import ffmpeg
from data.data_module import AVSRDataLoader
from tqdm import tqdm
from utils import save_vid_aud_txt, split_file
warnings.filterwarnings("ignore")
# Argument Parsing
parser = argparse.ArgumentParser(description="LRS3 Preprocessing")
parser.add_argument(
"--data-dir",
type=str,
help="The directory for sequence.",
)
parser.add_argument(
"--detector",
type=str,
default="retinaface",
help="Face detector used in the experiment.",
)
parser.add_argument(
"--dataset",
type=str,
help="Specify the dataset name used in the experiment",
)
parser.add_argument(
"--root-dir",
type=str,
help="The root directory of cropped-face dataset.",
)
parser.add_argument(
"--subset",
type=str,
required=True,
help="Subset of the dataset used in the experiment.",
)
parser.add_argument(
"--seg-duration",
type=int,
default=16,
help="Length of the segment in seconds.",
)
parser.add_argument(
"--groups",
type=int,
default=1,
help="Number of threads to be used in parallel.",
)
parser.add_argument(
"--job-index",
type=int,
default=0,
help="Index to identify separate jobs (useful for parallel processing).",
)
args = parser.parse_args()
seg_duration = args.seg_duration
dataset = args.dataset
args.data_dir = os.path.normpath(args.data_dir)
vid_dataloader = AVSRDataLoader(modality="video", detector=args.detector, resize=(96, 96))
aud_dataloader = AVSRDataLoader(modality="audio")
# Step 2, extract mouth patches from segments.
seg_vid_len = seg_duration * 25
seg_aud_len = seg_duration * 16000
label_filename = os.path.join(
args.root_dir,
"labels",
f"{dataset}_{args.subset}_transcript_lengths_seg{seg_duration}s.csv"
if args.groups <= 1
else f"{dataset}_{args.subset}_transcript_lengths_seg{seg_duration}s.{args.groups}.{args.job_index}.csv",
)
os.makedirs(os.path.dirname(label_filename), exist_ok=True)
print(f"Directory {os.path.dirname(label_filename)} created")
f = open(label_filename, "w")
# Step 2, extract mouth patches from segments.
dst_vid_dir = os.path.join(args.root_dir, dataset, dataset + f"_video_seg{seg_duration}s")
dst_txt_dir = os.path.join(args.root_dir, dataset, dataset + f"_text_seg{seg_duration}s")
if args.subset == "test":
filenames = glob.glob(os.path.join(args.data_dir, args.subset, "**", "*.mp4"), recursive=True)
elif args.subset == "train":
filenames = glob.glob(os.path.join(args.data_dir, "trainval", "**", "*.mp4"), recursive=True)
filenames.extend(glob.glob(os.path.join(args.data_dir, "pretrain", "**", "*.mp4"), recursive=True))
filenames.sort()
else:
raise NotImplementedError
unit = math.ceil(len(filenames) * 1.0 / args.groups)
filenames = filenames[args.job_index * unit : (args.job_index + 1) * unit]
for data_filename in tqdm(filenames):
try:
video_data = vid_dataloader.load_data(data_filename)
audio_data = aud_dataloader.load_data(data_filename)
except UnboundLocalError:
continue
if os.path.normpath(data_filename).split(os.sep)[-3] in ["trainval", "test"]:
dst_vid_filename = f"{data_filename.replace(args.data_dir, dst_vid_dir)[:-4]}.mp4"
dst_aud_filename = f"{data_filename.replace(args.data_dir, dst_vid_dir)[:-4]}.wav"
dst_txt_filename = f"{data_filename.replace(args.data_dir, dst_txt_dir)[:-4]}.txt"
trim_vid_data, trim_aud_data = video_data, audio_data
text_line_list = open(data_filename[:-4] + ".txt", "r").read().splitlines()[0].split(" ")
text_line = " ".join(text_line_list[2:])
content = text_line.replace("}", "").replace("{", "")
if trim_vid_data is None or trim_aud_data is None:
continue
video_length = len(trim_vid_data)
audio_length = trim_aud_data.size(1)
if video_length == 0 or audio_length == 0:
continue
if audio_length / video_length < 560.0 or audio_length / video_length > 720.0 or video_length < 12:
continue
save_vid_aud_txt(
dst_vid_filename,
dst_aud_filename,
dst_txt_filename,
trim_vid_data,
trim_aud_data,
content,
video_fps=25,
audio_sample_rate=16000,
)
in1 = ffmpeg.input(dst_vid_filename)
in2 = ffmpeg.input(dst_aud_filename)
out = ffmpeg.output(
in1["v"],
in2["a"],
dst_vid_filename[:-4] + ".m.mp4",
vcodec="copy",
acodec="aac",
strict="experimental",
loglevel="panic",
)
out.run()
os.remove(dst_aud_filename)
os.remove(dst_vid_filename)
shutil.move(dst_vid_filename[:-4] + ".m.mp4", dst_vid_filename)
basename = os.path.relpath(dst_vid_filename, start=os.path.join(args.root_dir, dataset))
f.write("{}\n".format(f"{dataset},{basename},{trim_vid_data.shape[0]},{len(content)}"))
continue
splitted = split_file(data_filename[:-4] + ".txt", max_frames=seg_vid_len)
for i in range(len(splitted)):
if len(splitted) == 1:
content, start, end, duration = splitted[i]
trim_vid_data, trim_aud_data = video_data, audio_data
else:
content, start, end, duration = splitted[i]
start_idx, end_idx = int(start * 25), int(end * 25)
try:
trim_vid_data, trim_aud_data = (
video_data[start_idx:end_idx],
audio_data[:, start_idx * 640 : end_idx * 640],
)
except TypeError:
continue
dst_vid_filename = f"{data_filename.replace(args.data_dir, dst_vid_dir)[:-4]}_{i:02d}.mp4"
dst_aud_filename = f"{data_filename.replace(args.data_dir, dst_vid_dir)[:-4]}_{i:02d}.wav"
dst_txt_filename = f"{data_filename.replace(args.data_dir, dst_txt_dir)[:-4]}_{i:02d}.txt"
if trim_vid_data is None or trim_aud_data is None:
continue
video_length = len(trim_vid_data)
audio_length = trim_aud_data.size(1)
if video_length == 0 or audio_length == 0:
continue
if audio_length / video_length < 560.0 or audio_length / video_length > 720.0 or video_length < 12:
continue
save_vid_aud_txt(
dst_vid_filename,
dst_aud_filename,
dst_txt_filename,
trim_vid_data,
trim_aud_data,
content,
video_fps=25,
audio_sample_rate=16000,
)
in1 = ffmpeg.input(dst_vid_filename)
in2 = ffmpeg.input(dst_aud_filename)
out = ffmpeg.output(
in1["v"],
in2["a"],
dst_vid_filename[:-4] + ".m.mp4",
vcodec="copy",
acodec="aac",
strict="experimental",
loglevel="panic",
)
out.run()
os.remove(dst_aud_filename)
os.remove(dst_vid_filename)
shutil.move(dst_vid_filename[:-4] + ".m.mp4", dst_vid_filename)
basename = os.path.relpath(dst_vid_filename, start=os.path.join(args.root_dir, dataset))
f.write("{}\n".format(f"{dataset},{basename},{trim_vid_data.shape[0]},{len(content)}"))
f.close()
|
