File size: 12,398 Bytes
4b3a024 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 |
import json
import logging
import os
import shutil
import subprocess
import tempfile
from dataclasses import dataclass
from glob import glob
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple, Union
import cv2
import ffmpeg
import numpy as np
import torch
from scipy import signal
from scipy.interpolate import interp1d
from scenedetect import ContentDetector, SceneManager, StatsManager
from scenedetect.video_manager import VideoManager
from detectors.s3fd import S3FD
from detectors.s3fd.nets import S3FDNet
from SyncNetInstance import SyncNetInstance
from SyncNetModel import S
# ---------------------------------------------------------------------- #
# Configuration #
# ---------------------------------------------------------------------- #
@dataclass
class PipelineConfig:
# Face-detection / tracking
facedet_scale: float = 0.25
crop_scale: float = 0.40
min_track: int = 50
frame_rate: int = 25
num_failed_det: int = 25
min_face_size: int = 100
# SyncNet
batch_size: int = 20
vshift: int = 15
# Local weight paths
s3fd_weights: str = "sfd_face.pth"
syncnet_weights: str = "syncnet_v2.model"
# Tools
ffmpeg_bin: str = "ffmpeg" # assumes ffmpeg in $PATH
audio_sample_rate: int = 16000 # resample rate for speech
@classmethod
def from_dict(cls, d: Dict[str, Any]):
return cls(**{k: v for k, v in d.items() if k in cls.__annotations__})
# ---------------------------------------------------------------------- #
# Pipeline #
# ---------------------------------------------------------------------- #
class SyncNetPipeline:
def __init__(
self,
cfg: Union[PipelineConfig, Dict[str, Any], None] = None,
*,
device: str = "cuda",
**override,
):
base = cfg if isinstance(cfg, PipelineConfig) else PipelineConfig.from_dict(cfg or {})
for k, v in override.items():
if hasattr(base, k):
setattr(base, k, v)
self.cfg = base
self.device = device
self.s3fd = self._load_s3fd(self.cfg.s3fd_weights)
self.syncnet = self._load_syncnet(self.cfg.syncnet_weights)
# ---------------------------- model loading ---------------------------- #
def _load_s3fd(self, path: str) -> S3FD:
logging.info(f"Loading S3FD from {path}")
net = S3FDNet(device=self.device)
net.load_state_dict(torch.load(path, map_location=self.device))
net.eval()
return S3FD(net=net, device=self.device)
def _load_syncnet(self, path: str) -> SyncNetInstance:
logging.info(f"Loading SyncNet from {path}")
model = S()
model.load_state_dict(torch.load(path, map_location=self.device))
model.eval()
return SyncNetInstance(net=model, device=self.device)
# ---------------------------- helpers ---------------------------------- #
@staticmethod
def _iou(a, b):
xA, yA = max(a[0], b[0]), max(a[1], b[1])
xB, yB = min(a[2], b[2]), min(a[3], b[3])
inter = max(0, xB - xA) * max(0, yB - yA)
areaA = (a[2] - a[0]) * (a[3] - a[1])
areaB = (b[2] - b[0]) * (b[3] - b[1])
return inter / (areaA + areaB - inter + 1e-8)
def _track(self, dets, *, min_track_override: Optional[int] = None):
cfg = self.cfg
min_track = max(
1, min_track_override if min_track_override is not None else cfg.min_track
)
tracks = []
while True:
t = []
for faces in dets:
for f in faces:
if not t:
t.append(f)
faces.remove(f)
elif (
f["frame"] - t[-1]["frame"] <= cfg.num_failed_det
and self._iou(f["bbox"], t[-1]["bbox"]) > 0.5
):
t.append(f)
faces.remove(f)
continue
else:
break
if not t:
break
if len(t) >= min_track:
fr = np.array([d["frame"] for d in t])
bb = np.array([d["bbox"] for d in t])
full_f = np.arange(fr[0], fr[-1] + 1)
bb_i = np.stack([interp1d(fr, bb[:, i])(full_f) for i in range(4)], 1)
if max(
np.mean(bb_i[:, 2] - bb_i[:, 0]),
np.mean(bb_i[:, 3] - bb_i[:, 1]),
) > cfg.min_face_size:
tracks.append({"frame": full_f, "bbox": bb_i})
return tracks
def _crop(self, track, frames, audio_wav, base):
cfg = self.cfg
base.parent.mkdir(parents=True, exist_ok=True)
tmp_avi = f"{base}t.avi"
vw = cv2.VideoWriter(tmp_avi, cv2.VideoWriter_fourcc(*"XVID"), cfg.frame_rate, (224, 224))
s, x, y = [], [], []
for b in track["bbox"]:
s.append(max(b[3] - b[1], b[2] - b[0]) / 2)
x.append((b[0] + b[2]) / 2)
y.append((b[1] + b[3]) / 2)
s, x, y = map(lambda v: signal.medfilt(v, 13), (s, x, y))
for i, fidx in enumerate(track["frame"]):
img = cv2.imread(frames[fidx])
if img is None:
continue
bs = s[i]
cs = cfg.crop_scale
pad = int(bs * (1 + 2 * cs))
img_p = cv2.copyMakeBorder(
img, pad, pad, pad, pad, cv2.BORDER_CONSTANT, value=(110, 110, 110)
)
my, mx = y[i] + pad, x[i] + pad
y1, y2 = int(my - bs), int(my + bs * (1 + 2 * cs))
x1, x2 = int(mx - bs * (1 + cs)), int(mx + bs * (1 + cs))
crop = cv2.resize(img_p[y1:y2, x1:x2], (224, 224))
vw.write(crop)
vw.release()
slice_wav = f"{base}.wav"
ss = track["frame"][0] / cfg.frame_rate
to = (track["frame"][-1] + 1) / cfg.frame_rate
subprocess.call(
f'{cfg.ffmpeg_bin} -y -i "{audio_wav}" -ss {ss:.3f} -to {to:.3f} "{slice_wav}"',
shell=True,
)
final_avi = f"{base}.avi"
subprocess.call(
f'{cfg.ffmpeg_bin} -y -i "{tmp_avi}" -i "{slice_wav}" -c:v copy -c:a copy "{final_avi}"',
shell=True,
)
os.remove(tmp_avi)
return final_avi
# ---------------------------- inference -------------------------------- #
def inference(
self,
video_path: str, # We do not extract audio from video_path!
audio_path: str,
*,
cache_dir: Optional[str] = None,
) -> Tuple[List[int], List[float], List[float], float, float, str, bool]:
cfg = self.cfg
work = Path(cache_dir) if cache_dir else Path(tempfile.mkdtemp())
if cache_dir:
work.mkdir(parents=True, exist_ok=True)
try:
# 1) Convert video to constant-fps AVI
avi = work / "video.avi"
(
ffmpeg.input(video_path)
.output(str(avi), **{"q:v": 2}, r=cfg.frame_rate, **{"async": 1})
.overwrite_output()
.run()
)
# 2) Extract frames
frames_dir = work / "frames"
frames_dir.mkdir(exist_ok=True)
(
ffmpeg.input(str(avi))
.output(str(frames_dir / "%06d.jpg"), **{"q:v": 2}, f="image2", threads=1)
.overwrite_output()
.run()
)
frames = sorted(glob(str(frames_dir / "*.jpg")))
# 3) Resample speech
audio_wav = work / "speech.wav"
(
ffmpeg.input(audio_path)
.output(str(audio_wav), ac=1, ar=cfg.audio_sample_rate, format="wav")
.overwrite_output()
.run()
)
# 4) Face detection
detections = []
for i, fp in enumerate(frames):
img = cv2.imread(fp)
boxes = (
self.s3fd.detect_faces(
cv2.cvtColor(img, cv2.COLOR_BGR2RGB),
conf_th=0.9,
scales=[cfg.facedet_scale],
)
if img is not None
else []
)
detections.append(
[
{"frame": i, "bbox": b[:-1].tolist(), "conf": float(b[-1])}
for b in boxes
]
)
flat = [f for fs in detections for f in fs]
s3fd_json = json.dumps(flat) if flat else ""
has_face = bool(flat)
# 5) Scene detection
vm = VideoManager([str(avi)])
sm = SceneManager(StatsManager())
sm.add_detector(ContentDetector())
vm.start()
sm.detect_scenes(frame_source=vm)
scenes = sm.get_scene_list(vm.get_base_timecode()) or [
(vm.get_base_timecode(), vm.get_current_timecode())
]
total_frames = len(detections)
# 6) Track faces
tracks = []
for sc in scenes:
s = min(max(sc[0].frame_num, 0), total_frames)
end_tc = sc[1] if sc[1] is not None else vm.get_current_timecode()
e = min(max(end_tc.frame_num, s), total_frames)
if e <= s:
continue
scene_detections = [lst.copy() for lst in detections[s:e]]
if not scene_detections:
continue
min_track_len = max(1, min(cfg.min_track, len(scene_detections)))
tracks.extend(
self._track(
scene_detections,
min_track_override=min_track_len,
)
)
# 7) Crop tracks
crops = [
self._crop(t, frames, str(audio_wav), Path(work) / "cropped" / f"{i:05d}") for i, t in enumerate(tracks)
]
# AV offset: 5
# Min dist: 5.370
# Confidence: 9.892
# crops = [work / ".." / ".."/ "data" / "example.avi"]
# AV offset: 3
# Min dist: 5.348
# Confidence: 10.081
# crops = [work / "video.avi"]
# AV offset: 3
# Min dist: 6.668
# Confidence: 8.337
# 8) SyncNet evaluation
offsets, confs, dists = [], [], []
class Opt: ...
for i, cp in enumerate(crops):
crop_dir = work / "cropped" / f"crop_{i:05d}"
frames_dir = crop_dir
frames_dir.mkdir(parents=True, exist_ok=True)
audio_path = crop_dir / "audio.wav"
# Extract frames
(
ffmpeg.input(cp)
.output(str(frames_dir / "%06d.jpg"), f="image2", threads=1)
.overwrite_output()
.run()
)
# Extract audio
(
ffmpeg.input(cp)
.output(
str(audio_path),
ac=1,
vn=None,
acodec="pcm_s16le",
ar=16000,
af="aresample=async=1",
)
.overwrite_output()
.run()
)
opt = Opt()
opt.tmp_dir = str(crop_dir)
opt.batch_size = cfg.batch_size
opt.vshift = cfg.vshift
off, conf, dist = self.syncnet.evaluate(opt=opt)
offsets.append(off)
confs.append(conf)
dists.append(dist)
if not offsets:
return ([], [], [], 0.0, 0.0, "", False)
return offsets, confs, dists, max(confs), min(dists), s3fd_json, has_face
finally:
if not cache_dir:
shutil.rmtree(work, ignore_errors=True)
|