Update emotion.py
Browse files- emotion.py +21 -29
emotion.py
CHANGED
|
@@ -24,14 +24,12 @@ SWITCH_CONFIRM_SEC = 2.5
|
|
| 24 |
IOU_TH = 0.25
|
| 25 |
ZOOM_FACTOR = 1.5
|
| 26 |
|
| 27 |
-
DETECTOR_BACKEND = "mtcnn"
|
| 28 |
ALIGN_FACE = True
|
| 29 |
|
| 30 |
-
# ✅
|
| 31 |
-
# 臉框面積佔整張畫面比例下限(
|
| 32 |
-
|
| 33 |
-
# 偵測器信心值下限(MTCNN/RetinaFace 通常有 confidence)
|
| 34 |
-
MIN_DET_CONF = 0.90 # 可調:0.80~0.95
|
| 35 |
|
| 36 |
|
| 37 |
# =========================
|
|
@@ -59,9 +57,6 @@ EMO_TO_SCORE = {
|
|
| 59 |
"unknown": 3,
|
| 60 |
}
|
| 61 |
|
| 62 |
-
def get_emotion_score(emo: str) -> int:
|
| 63 |
-
return EMO_TO_SCORE.get(emo, 3)
|
| 64 |
-
|
| 65 |
SCORE_LABELS = {
|
| 66 |
1: "1(心情差)",
|
| 67 |
2: "2(不太好)",
|
|
@@ -83,7 +78,7 @@ class AppState:
|
|
| 83 |
finished: bool = False
|
| 84 |
last_analyze_t: float = 0.0
|
| 85 |
|
| 86 |
-
# tracking
|
| 87 |
track_bbox: Optional[Tuple[int, int, int, int]] = None
|
| 88 |
|
| 89 |
# smoothing
|
|
@@ -98,10 +93,9 @@ class AppState:
|
|
| 98 |
# final
|
| 99 |
final_emo: Optional[str] = None
|
| 100 |
final_conf: Optional[float] = None
|
| 101 |
-
|
| 102 |
final_score: int = 3
|
| 103 |
|
| 104 |
-
# 最近有沒有臉(避免節流期間
|
| 105 |
last_face_t: float = 0.0
|
| 106 |
face_in_box: bool = False
|
| 107 |
|
|
@@ -313,14 +307,13 @@ def _reset_state_for_camera(st: AppState) -> AppState:
|
|
| 313 |
st.final_emo = None
|
| 314 |
st.final_conf = None
|
| 315 |
st.final_score = 3
|
| 316 |
-
|
| 317 |
st.last_face_t = 0.0
|
| 318 |
st.face_in_box = False
|
| 319 |
return st
|
| 320 |
|
| 321 |
|
| 322 |
# =========================
|
| 323 |
-
# ✅ 判斷
|
| 324 |
# =========================
|
| 325 |
def _is_real_face(
|
| 326 |
pick: Dict,
|
|
@@ -330,13 +323,10 @@ def _is_real_face(
|
|
| 330 |
fh, fw = frame_hw
|
| 331 |
x, y, w, h = bbox_full_xywh
|
| 332 |
|
| 333 |
-
# 1) 面積門檻(避免小塊背景誤判)
|
| 334 |
area_ratio = (w * h) / max(1.0, float(fw * fh))
|
| 335 |
if area_ratio < MIN_FACE_AREA_RATIO:
|
| 336 |
return False
|
| 337 |
|
| 338 |
-
# 2) 偵測 confidence(如果偵測器有提供)
|
| 339 |
-
# DeepFace.extract_faces 的 dict 常見會有 "confidence"(依 backend 而異)
|
| 340 |
det_conf = pick.get("confidence", None)
|
| 341 |
if det_conf is not None:
|
| 342 |
try:
|
|
@@ -352,9 +342,9 @@ def _is_real_face(
|
|
| 352 |
# Camera streaming
|
| 353 |
# =========================
|
| 354 |
def on_stream(frame_rgb: np.ndarray, st: AppState):
|
|
|
|
| 355 |
if st.finished or (not st.running):
|
| 356 |
cam_visible = False
|
| 357 |
-
|
| 358 |
if st.final_emo:
|
| 359 |
result_html = _result_card_html(st.final_emo, st.final_conf)
|
| 360 |
mood_val = _score_radio_value(EMO_TO_SCORE.get(st.final_emo, 3))
|
|
@@ -387,7 +377,7 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
|
|
| 387 |
|
| 388 |
now = time.time()
|
| 389 |
|
| 390 |
-
# 節流:沿用最近一次
|
| 391 |
detected_in_box_recent = st.face_in_box and (now - st.last_face_t) < 1.0
|
| 392 |
|
| 393 |
if now - st.last_analyze_t < ANALYZE_EVERY_SEC:
|
|
@@ -406,12 +396,12 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
|
|
| 406 |
|
| 407 |
small = _downsample_rgb(frame_rgb, DOWNSAMPLE_W)
|
| 408 |
|
| 409 |
-
# ✅
|
| 410 |
try:
|
| 411 |
faces = DeepFace.extract_faces(
|
| 412 |
img_path=small,
|
| 413 |
detector_backend=DETECTOR_BACKEND,
|
| 414 |
-
enforce_detection=True,
|
| 415 |
align=ALIGN_FACE,
|
| 416 |
)
|
| 417 |
except Exception:
|
|
@@ -419,7 +409,7 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
|
|
| 419 |
|
| 420 |
pick, bbox = _pick_face_by_tracking(faces, st.track_bbox)
|
| 421 |
|
| 422 |
-
# (A) 完全沒臉
|
| 423 |
if pick is None or bbox is None or pick.get("face") is None:
|
| 424 |
st.track_bbox = None
|
| 425 |
st.face_in_box = False
|
|
@@ -445,7 +435,7 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
|
|
| 445 |
x, y, w, h = bbox
|
| 446 |
bbox_full = (x * sx, y * sy, w * sx, h * sy)
|
| 447 |
|
| 448 |
-
# ✅
|
| 449 |
if not _is_real_face(pick, bbox_full, (fh, fw)):
|
| 450 |
st.track_bbox = None
|
| 451 |
st.face_in_box = False
|
|
@@ -469,7 +459,7 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
|
|
| 469 |
st.face_in_box = detected_in_box
|
| 470 |
st.track_bbox = bbox
|
| 471 |
|
| 472 |
-
# (B) 有臉但不在框內
|
| 473 |
if not detected_in_box:
|
| 474 |
annotated = _draw_guide_overlay(frame_rgb, False)
|
| 475 |
return (
|
|
@@ -481,12 +471,12 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
|
|
| 481 |
gr.update(value="重新攝影機辨���"),
|
| 482 |
)
|
| 483 |
|
| 484 |
-
# (C) 有臉且在框內:綠框,
|
| 485 |
face = pick["face"]
|
| 486 |
if face.dtype != np.uint8:
|
| 487 |
face = np.clip(face * 255.0, 0, 255).astype(np.uint8)
|
| 488 |
|
| 489 |
-
# ✅ analyze 也
|
| 490 |
try:
|
| 491 |
result = DeepFace.analyze(
|
| 492 |
img_path=face,
|
|
@@ -551,6 +541,7 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
|
|
| 551 |
stable_key = st.emo_keys[stable_idx]
|
| 552 |
st.stable_emo = stable_key
|
| 553 |
|
|
|
|
| 554 |
if st.candidate_emo != stable_key:
|
| 555 |
st.candidate_emo = stable_key
|
| 556 |
st.candidate_since = now
|
|
@@ -597,7 +588,7 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
|
|
| 597 |
def on_restart(st: AppState):
|
| 598 |
st = _reset_state_for_camera(st)
|
| 599 |
return (
|
| 600 |
-
gr.update(visible=True),
|
| 601 |
gr.update(value=_hint_html("等待影像…(尚未開始串流)")),
|
| 602 |
gr.update(value=_score_radio_value(3)),
|
| 603 |
gr.update(value=_camera_hint_before_start()),
|
|
@@ -608,9 +599,10 @@ def on_restart(st: AppState):
|
|
| 608 |
def on_stop(st: AppState):
|
| 609 |
st.running = False
|
| 610 |
st.finished = True
|
|
|
|
| 611 |
if st.final_emo is None:
|
| 612 |
return (
|
| 613 |
-
gr.update(visible=False),
|
| 614 |
gr.update(value=_hint_html("已停止辨識。")),
|
| 615 |
gr.update(value=_score_radio_value(3)),
|
| 616 |
gr.update(value=_camera_hint_stopped()),
|
|
@@ -619,7 +611,7 @@ def on_stop(st: AppState):
|
|
| 619 |
)
|
| 620 |
|
| 621 |
return (
|
| 622 |
-
gr.update(visible=False),
|
| 623 |
gr.update(value=_result_card_html(st.final_emo, st.final_conf)),
|
| 624 |
gr.update(value=_score_radio_value(EMO_TO_SCORE.get(st.final_emo, 3))),
|
| 625 |
gr.update(value=_camera_hint_done()),
|
|
|
|
| 24 |
IOU_TH = 0.25
|
| 25 |
ZOOM_FACTOR = 1.5
|
| 26 |
|
| 27 |
+
DETECTOR_BACKEND = "mtcnn" # 想快一點可改 "opencv"
|
| 28 |
ALIGN_FACE = True
|
| 29 |
|
| 30 |
+
# ✅ 防止背景誤判為臉(重要)
|
| 31 |
+
MIN_FACE_AREA_RATIO = 0.05 # 臉框面積佔整張畫面比例下限(3%~8%可調)
|
| 32 |
+
MIN_DET_CONF = 0.90 # 偵測器信心值下限(若 backend 提供 confidence)
|
|
|
|
|
|
|
| 33 |
|
| 34 |
|
| 35 |
# =========================
|
|
|
|
| 57 |
"unknown": 3,
|
| 58 |
}
|
| 59 |
|
|
|
|
|
|
|
|
|
|
| 60 |
SCORE_LABELS = {
|
| 61 |
1: "1(心情差)",
|
| 62 |
2: "2(不太好)",
|
|
|
|
| 78 |
finished: bool = False
|
| 79 |
last_analyze_t: float = 0.0
|
| 80 |
|
| 81 |
+
# tracking(用於連續幀追蹤)
|
| 82 |
track_bbox: Optional[Tuple[int, int, int, int]] = None
|
| 83 |
|
| 84 |
# smoothing
|
|
|
|
| 93 |
# final
|
| 94 |
final_emo: Optional[str] = None
|
| 95 |
final_conf: Optional[float] = None
|
|
|
|
| 96 |
final_score: int = 3
|
| 97 |
|
| 98 |
+
# ✅ 新增:最近有沒有臉(避免節流期間框顏色亂跳)
|
| 99 |
last_face_t: float = 0.0
|
| 100 |
face_in_box: bool = False
|
| 101 |
|
|
|
|
| 307 |
st.final_emo = None
|
| 308 |
st.final_conf = None
|
| 309 |
st.final_score = 3
|
|
|
|
| 310 |
st.last_face_t = 0.0
|
| 311 |
st.face_in_box = False
|
| 312 |
return st
|
| 313 |
|
| 314 |
|
| 315 |
# =========================
|
| 316 |
+
# ✅ 判斷這個 bbox 是否真的像「人臉」
|
| 317 |
# =========================
|
| 318 |
def _is_real_face(
|
| 319 |
pick: Dict,
|
|
|
|
| 323 |
fh, fw = frame_hw
|
| 324 |
x, y, w, h = bbox_full_xywh
|
| 325 |
|
|
|
|
| 326 |
area_ratio = (w * h) / max(1.0, float(fw * fh))
|
| 327 |
if area_ratio < MIN_FACE_AREA_RATIO:
|
| 328 |
return False
|
| 329 |
|
|
|
|
|
|
|
| 330 |
det_conf = pick.get("confidence", None)
|
| 331 |
if det_conf is not None:
|
| 332 |
try:
|
|
|
|
| 342 |
# Camera streaming
|
| 343 |
# =========================
|
| 344 |
def on_stream(frame_rgb: np.ndarray, st: AppState):
|
| 345 |
+
# 已結束/停止
|
| 346 |
if st.finished or (not st.running):
|
| 347 |
cam_visible = False
|
|
|
|
| 348 |
if st.final_emo:
|
| 349 |
result_html = _result_card_html(st.final_emo, st.final_conf)
|
| 350 |
mood_val = _score_radio_value(EMO_TO_SCORE.get(st.final_emo, 3))
|
|
|
|
| 377 |
|
| 378 |
now = time.time()
|
| 379 |
|
| 380 |
+
# 節流:沿用最近一次臉狀態(避免顏色亂跳)
|
| 381 |
detected_in_box_recent = st.face_in_box and (now - st.last_face_t) < 1.0
|
| 382 |
|
| 383 |
if now - st.last_analyze_t < ANALYZE_EVERY_SEC:
|
|
|
|
| 396 |
|
| 397 |
small = _downsample_rgb(frame_rgb, DOWNSAMPLE_W)
|
| 398 |
|
| 399 |
+
# ✅ extract_faces 變嚴格:沒臉就當 faces=[]
|
| 400 |
try:
|
| 401 |
faces = DeepFace.extract_faces(
|
| 402 |
img_path=small,
|
| 403 |
detector_backend=DETECTOR_BACKEND,
|
| 404 |
+
enforce_detection=True,
|
| 405 |
align=ALIGN_FACE,
|
| 406 |
)
|
| 407 |
except Exception:
|
|
|
|
| 409 |
|
| 410 |
pick, bbox = _pick_face_by_tracking(faces, st.track_bbox)
|
| 411 |
|
| 412 |
+
# (A) 完全沒臉
|
| 413 |
if pick is None or bbox is None or pick.get("face") is None:
|
| 414 |
st.track_bbox = None
|
| 415 |
st.face_in_box = False
|
|
|
|
| 435 |
x, y, w, h = bbox
|
| 436 |
bbox_full = (x * sx, y * sy, w * sx, h * sy)
|
| 437 |
|
| 438 |
+
# ✅ 防背景誤判:不夠大/信心不足 -> 視為沒臉
|
| 439 |
if not _is_real_face(pick, bbox_full, (fh, fw)):
|
| 440 |
st.track_bbox = None
|
| 441 |
st.face_in_box = False
|
|
|
|
| 459 |
st.face_in_box = detected_in_box
|
| 460 |
st.track_bbox = bbox
|
| 461 |
|
| 462 |
+
# (B) 有臉但不在框內
|
| 463 |
if not detected_in_box:
|
| 464 |
annotated = _draw_guide_overlay(frame_rgb, False)
|
| 465 |
return (
|
|
|
|
| 471 |
gr.update(value="重新攝影機辨���"),
|
| 472 |
)
|
| 473 |
|
| 474 |
+
# (C) 有臉且在框內:綠框,開始 analyze
|
| 475 |
face = pick["face"]
|
| 476 |
if face.dtype != np.uint8:
|
| 477 |
face = np.clip(face * 255.0, 0, 255).astype(np.uint8)
|
| 478 |
|
| 479 |
+
# ✅ analyze 也嚴格:避免誤判一路走到底
|
| 480 |
try:
|
| 481 |
result = DeepFace.analyze(
|
| 482 |
img_path=face,
|
|
|
|
| 541 |
stable_key = st.emo_keys[stable_idx]
|
| 542 |
st.stable_emo = stable_key
|
| 543 |
|
| 544 |
+
# 延遲確認
|
| 545 |
if st.candidate_emo != stable_key:
|
| 546 |
st.candidate_emo = stable_key
|
| 547 |
st.candidate_since = now
|
|
|
|
| 588 |
def on_restart(st: AppState):
|
| 589 |
st = _reset_state_for_camera(st)
|
| 590 |
return (
|
| 591 |
+
gr.update(visible=True, value=None), # ✅ 清掉上一張畫面(你問的重點)
|
| 592 |
gr.update(value=_hint_html("等待影像…(尚未開始串流)")),
|
| 593 |
gr.update(value=_score_radio_value(3)),
|
| 594 |
gr.update(value=_camera_hint_before_start()),
|
|
|
|
| 599 |
def on_stop(st: AppState):
|
| 600 |
st.running = False
|
| 601 |
st.finished = True
|
| 602 |
+
|
| 603 |
if st.final_emo is None:
|
| 604 |
return (
|
| 605 |
+
gr.update(visible=False, value=None), # ✅ 順便清空畫面(可選,但我先幫你加上)
|
| 606 |
gr.update(value=_hint_html("已停止辨識。")),
|
| 607 |
gr.update(value=_score_radio_value(3)),
|
| 608 |
gr.update(value=_camera_hint_stopped()),
|
|
|
|
| 611 |
)
|
| 612 |
|
| 613 |
return (
|
| 614 |
+
gr.update(visible=False, value=None), # ✅ 順便清空
|
| 615 |
gr.update(value=_result_card_html(st.final_emo, st.final_conf)),
|
| 616 |
gr.update(value=_score_radio_value(EMO_TO_SCORE.get(st.final_emo, 3))),
|
| 617 |
gr.update(value=_camera_hint_done()),
|