Spaces:
Sleeping
Sleeping
File size: 14,047 Bytes
5029d29 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 | """Sign recogniser: turn a stream of landmark frames into LSF glosses.
There is no off-the-shelf, production-grade LSF translation model, and training
one needs a large annotated corpus. So this recogniser is built to be *useful
out of the box and extensible*:
1. **Motion segmentation** — watches hand presence + motion energy to find where
one sign begins and ends (so we classify gestures, not every frame).
2. **Learn-by-example (DTW)** — the primary path. You record a few samples of
each sign through the UI; we store the normalised landmark sequence as a
template and match new segments with Dynamic Time Warping. This works on a
laptop CPU, with no GPU and no training run, for a small/medium vocabulary.
3. **Deep-model hook** — if a trained Keras model (`model.h5`) and its label map
are present, we use it instead of/along DTW. Plug your own sequence model in.
4. **Heuristic fallbacks** — a couple of signs detectable by simple geometry so
the demo produces output before you've recorded anything.
Everything operates on the body-relative feature vectors from `features.py`.
"""
from __future__ import annotations
import json
import os
import time
from dataclasses import dataclass, field
from typing import List, Optional
import numpy as np
from .features import (
FEATURE_DIM,
N_FACE_SEL,
N_POSE_SEL,
extract_features,
mouth_openness,
)
from .landmarks import LandmarkFrame
# Offsets into the feature vector (see features.extract_features layout).
_POSE_LEN = N_POSE_SEL * 3
_FACE_LEN = N_FACE_SEL * 3
_HANDS_LEN = 21 * 3 * 2
_HANDS_START = _POSE_LEN + _FACE_LEN
# Offsets *within the matching vector* ([pose(27), left hand(63), right hand(63)]).
# POSE_IDX order is [nose, L/R shoulder, L/R elbow, L/R wrist, L/R hip]; the arm
# block (elbows+wrists, selection indices 3..6) is coords 9..21.
_ARM_SLICE = slice(9, 21)
_LEFT_HAND_SLICE = slice(_POSE_LEN, _POSE_LEN + 63)
_RIGHT_HAND_SLICE = slice(_POSE_LEN + 63, _POSE_LEN + 126)
# Tunables
TEMPLATE_LEN = 32 # frames each segment is resampled to before DTW
MIN_SEGMENT_FRAMES = 6 # ignore micro-twitches
MOTION_THRESHOLD = 0.025 # per-frame RMS motion (of moving parts) to be "active"
SETTLE_FRAMES = 8 # consecutive calm frames that close a segment
DTW_BAND = 8 # Sakoe-Chiba band radius
REJECT_DISTANCE = 3.5 # avg aligned distance above which we emit nothing
@dataclass
class Recognition:
"""A recognised sign."""
gloss: str
confidence: float
timestamp: float = field(default_factory=time.time)
source: str = "dtw" # "dtw" | "model" | "heuristic"
def _matching_vector(feat: np.ndarray) -> np.ndarray:
"""Pose + hands slice of a feature vector (face/flags dropped for matching).
Facial expression varies a lot between repetitions of the same sign, so we
align on the manual channel (arms + both hands) which carries the gesture.
"""
pose = feat[:_POSE_LEN]
hands = feat[_HANDS_START:_HANDS_START + _HANDS_LEN]
return np.concatenate([pose, hands])
def _resample(seq: np.ndarray, n: int) -> np.ndarray:
"""Linearly resample a (T, D) sequence to (n, D) along the time axis."""
t = seq.shape[0]
if t == n:
return seq
if t == 1:
return np.repeat(seq, n, axis=0)
src = np.linspace(0.0, 1.0, t)
dst = np.linspace(0.0, 1.0, n)
out = np.empty((n, seq.shape[1]), dtype=np.float32)
for d in range(seq.shape[1]):
out[:, d] = np.interp(dst, src, seq[:, d])
return out
def _dtw_distance(a: np.ndarray, b: np.ndarray, band: int = DTW_BAND) -> float:
"""Banded DTW distance between two (N, D) sequences, normalised by path len."""
n, m = a.shape[0], b.shape[0]
inf = np.inf
cost = np.full((n + 1, m + 1), inf, dtype=np.float64)
cost[0, 0] = 0.0
for i in range(1, n + 1):
j_lo = max(1, i - band)
j_hi = min(m, i + band)
ai = a[i - 1]
for j in range(j_lo, j_hi + 1):
d = float(np.linalg.norm(ai - b[j - 1]))
cost[i, j] = d + min(cost[i - 1, j], cost[i, j - 1], cost[i - 1, j - 1])
# Normalise by the diagonal path length so different lengths compare fairly.
return cost[n, m] / (n + m)
class SignRecognizer:
"""Stateful streaming recogniser.
Call `update(frame, t)` once per video frame. It returns a `Recognition`
on the frame where a completed sign is identified, otherwise None.
"""
def __init__(
self,
templates_path: str = "signs/lsf_signs.json",
model_path: Optional[str] = None,
*,
use_heuristics: bool = True,
) -> None:
self.templates_path = templates_path
self.use_heuristics = use_heuristics
# templates: gloss -> list of (TEMPLATE_LEN, D) matching sequences
self.templates: dict[str, List[np.ndarray]] = {}
self._load_templates()
# streaming state
self._buffer: List[np.ndarray] = [] # matching vectors in segment
self._prev_match: Optional[np.ndarray] = None
self._calm_run = 0
self._active = False
self._last_emit = 0.0
# recording state (learn-by-example)
self._recording = False
self._record_buffer: List[np.ndarray] = []
self._record_label: Optional[str] = None
# optional deep model
self._model = None
self._labels: List[str] = []
if model_path and os.path.exists(model_path):
self._try_load_model(model_path)
# -- persistence ----------------------------------------------------------
def _load_templates(self) -> None:
if not os.path.exists(self.templates_path):
return
with open(self.templates_path, "r", encoding="utf-8") as fh:
raw = json.load(fh)
for gloss, samples in raw.items():
self.templates[gloss] = [np.asarray(s, dtype=np.float32) for s in samples]
def _save_templates(self) -> None:
serialisable = {
gloss: [s.tolist() for s in samples]
for gloss, samples in self.templates.items()
}
try:
os.makedirs(os.path.dirname(self.templates_path) or ".", exist_ok=True)
with open(self.templates_path, "w", encoding="utf-8") as fh:
json.dump(serialisable, fh)
except OSError:
# Read-only/ephemeral filesystem (e.g. a hosted Space): keep the
# templates in memory so the session still works, just not persisted.
pass
@property
def vocabulary(self) -> List[str]:
return sorted(self.templates.keys())
def template_counts(self) -> dict[str, int]:
"""How many recorded samples back each trained gloss."""
return {gloss: len(samples) for gloss, samples in self.templates.items()}
def delete_gloss(self, gloss: str) -> bool:
"""Forget a trained sign entirely. Returns True if it existed."""
if gloss in self.templates:
del self.templates[gloss]
self._save_templates()
return True
return False
# -- deep model hook ------------------------------------------------------
def _try_load_model(self, model_path: str) -> None:
try:
import tensorflow as tf # noqa: F401
from tensorflow import keras
self._model = keras.models.load_model(model_path)
label_path = os.path.splitext(model_path)[0] + ".labels.json"
if os.path.exists(label_path):
with open(label_path, encoding="utf-8") as fh:
self._labels = json.load(fh)
except Exception:
self._model = None # stay on DTW/heuristics
# -- recording API (learn-by-example) ------------------------------------
def start_recording(self, label: str) -> None:
self._recording = True
self._record_label = label
self._record_buffer = []
def stop_recording(self) -> Optional[str]:
"""Finalise a recording into a template. Returns the label or None."""
self._recording = False
if not self._record_label or len(self._record_buffer) < MIN_SEGMENT_FRAMES:
self._record_buffer = []
return None
seq = _resample(np.asarray(self._record_buffer, dtype=np.float32), TEMPLATE_LEN)
self.templates.setdefault(self._record_label, []).append(seq)
self._save_templates()
label = self._record_label
self._record_label = None
self._record_buffer = []
return label
def cancel_recording(self) -> None:
"""Abort the current recording without saving anything."""
self._recording = False
self._record_label = None
self._record_buffer = []
@property
def is_recording(self) -> bool:
return self._recording
# -- motion gating --------------------------------------------------------
def _motion_energy(self, frame: LandmarkFrame, match: np.ndarray) -> float:
"""RMS per-coordinate displacement over the present moving body parts.
Uses the hands when visible (the primary signing channel), else falls
back to the arms (elbows/wrists), so segmentation works for one hand,
two hands, or arm-only gestures without static channels washing it out.
"""
if self._prev_match is None:
return 0.0
diff = match - self._prev_match
idx = []
if frame.left_hand is not None:
idx.append(diff[_LEFT_HAND_SLICE])
if frame.right_hand is not None:
idx.append(diff[_RIGHT_HAND_SLICE])
if not idx: # no hands -> use arm motion
idx.append(diff[_ARM_SLICE])
d = np.concatenate(idx)
return float(np.linalg.norm(d)) / np.sqrt(d.size)
# -- main streaming entry point ------------------------------------------
def update(self, frame: LandmarkFrame, t: Optional[float] = None) -> Optional[Recognition]:
t = time.time() if t is None else t
feat = extract_features(frame)
match = _matching_vector(feat)
if self._recording:
self._record_buffer.append(match)
# Motion energy between consecutive frames, measured only over the body
# parts that are actually present (so static/zero channels don't dilute it).
motion = self._motion_energy(frame, match)
self._prev_match = match
active_now = motion > MOTION_THRESHOLD
if active_now:
if not self._active:
self._active = True
self._buffer = []
self._buffer.append(match)
self._calm_run = 0
return None
# Not active this frame.
if self._active:
self._calm_run += 1
if self._calm_run >= SETTLE_FRAMES:
# Segment finished -> classify.
segment = self._buffer
self._active = False
self._buffer = []
self._calm_run = 0
if len(segment) >= MIN_SEGMENT_FRAMES and (t - self._last_emit) > 0.3:
rec = self._classify(np.asarray(segment, dtype=np.float32), frame, t)
if rec is not None:
self._last_emit = t
return rec
return None
# -- classification -------------------------------------------------------
def _classify(
self, segment: np.ndarray, frame: LandmarkFrame, t: float
) -> Optional[Recognition]:
seq = _resample(segment, TEMPLATE_LEN)
if self._model is not None:
rec = self._classify_model(seq, t)
if rec is not None:
return rec
rec = self._classify_dtw(seq, t)
if rec is not None:
return rec
if self.use_heuristics:
return self._classify_heuristic(segment, frame, t)
return None
def _classify_model(self, seq: np.ndarray, t: float) -> Optional[Recognition]:
probs = self._model.predict(seq[None, ...], verbose=0)[0]
idx = int(np.argmax(probs))
conf = float(probs[idx])
if conf < 0.5 or idx >= len(self._labels):
return None
return Recognition(self._labels[idx], conf, t, source="model")
def _classify_dtw(self, seq: np.ndarray, t: float) -> Optional[Recognition]:
if not self.templates:
return None
best_gloss, best_dist = None, np.inf
for gloss, samples in self.templates.items():
for tmpl in samples:
d = _dtw_distance(seq, tmpl)
if d < best_dist:
best_dist, best_gloss = d, gloss
if best_gloss is None or best_dist > REJECT_DISTANCE:
return None
conf = max(0.0, 1.0 - best_dist / REJECT_DISTANCE)
return Recognition(best_gloss, conf, t, source="dtw")
def _classify_heuristic(
self, segment: np.ndarray, frame: LandmarkFrame, t: float
) -> Optional[Recognition]:
"""A couple of geometry-only signs so the demo isn't silent.
These are deliberately simple and approximate — real coverage comes from
recorded templates or a trained model.
"""
# Lateral oscillation of an open hand -> greeting "Bonjour". We track the
# x of the first hand-wrist channel (right after the pose block).
if segment.shape[0] >= MIN_SEGMENT_FRAMES:
x_track = segment[:, _POSE_LEN] # x of left-hand wrist in matching vec
sweep = float(np.max(x_track) - np.min(x_track))
zero_cross = np.sum(np.abs(np.diff(np.sign(np.diff(x_track)))) > 0)
if sweep > 0.6 and zero_cross >= 2:
return Recognition("BONJOUR", 0.4, t, source="heuristic")
# Open mouth + raised hand held still -> tentative "OUI" placeholder.
if mouth_openness(frame) > 0.35 and frame.has_hands:
return Recognition("OUI", 0.3, t, source="heuristic")
return None
|