webolavo commited on
Commit
a2fefc0
·
verified ·
1 Parent(s): 34ffa87

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +84 -12
app.py CHANGED
@@ -27,7 +27,24 @@ PERSON_PHRASES = {
27
  "male": ("man", "male person"),
28
  "child": ("young child", "baby", "toddler", "little kid"),
29
  }
30
- PART_TYPES = ("face", "hand", "chest", "torso", "arm", "body", "leg", "thigh", "foot")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31
 
32
  PHRASE_SPECS: list[dict[str, str]] = []
33
  for phrase in PERSON_PHRASES["female"]:
@@ -71,6 +88,17 @@ CHILD_MARGIN = float(os.getenv("CHILD_MARGIN", "0.05"))
71
  # A body part must also show a meaningful female/male separation before it can
72
  # decide the image on its own. This does not add another model forward pass.
73
  PART_GENDER_MARGIN = float(os.getenv("PART_GENDER_MARGIN", "0.05"))
 
 
 
 
 
 
 
 
 
 
 
74
  NMS_IOU_THRESHOLD = float(os.getenv("NMS_IOU_THRESHOLD", "0.75"))
75
  MATCH_IOU_THRESHOLD = float(os.getenv("MATCH_IOU_THRESHOLD", "0.25"))
76
  MATCH_CONTAINMENT_THRESHOLD = float(os.getenv("MATCH_CONTAINMENT_THRESHOLD", "0.80"))
@@ -88,6 +116,9 @@ for threshold_name, threshold_value in {
88
  "GENDER_MARGIN": GENDER_MARGIN,
89
  "CHILD_MARGIN": CHILD_MARGIN,
90
  "PART_GENDER_MARGIN": PART_GENDER_MARGIN,
 
 
 
91
  "NMS_IOU_THRESHOLD": NMS_IOU_THRESHOLD,
92
  "MATCH_IOU_THRESHOLD": MATCH_IOU_THRESHOLD,
93
  "MATCH_CONTAINMENT_THRESHOLD": MATCH_CONTAINMENT_THRESHOLD,
@@ -517,35 +548,73 @@ def _run_grounding_dino(image: Image.Image) -> dict[str, Any]:
517
 
518
  for part in body_parts:
519
  owner = _best_matching_person(part["box"], persons)
520
- part_margin_is_clear = part["score_margin"] >= PART_GENDER_MARGIN
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
521
 
522
  if owner is not None:
523
  owner_hint = owner["winner"]
 
 
 
 
524
  part["owner_gender"] = owner["gender"]
525
  part["owner_winner"] = owner_hint
526
  part["owner_confirmed"] = owner["confirmed"]
527
  part["owner_score"] = owner["score"]
 
528
 
529
  if owner["confirmed"]:
530
  part["attribution"] = owner["gender"]
531
  part["attribution_source"] = f"confirmed_{owner['gender']}_person_box"
532
- elif owner_hint in {"male", "child"}:
533
  # An ambiguous whole-person result still provides stronger
534
- # context than a standalone gender guess for a hand or face.
 
535
  part["attribution"] = owner_hint
536
  part["attribution_source"] = f"ambiguous_{owner_hint}_person_hint"
537
- elif part["raw_gender"] == "female" and part_margin_is_clear:
538
- # The whole person and the body part both lean female.
 
539
  part["attribution"] = "female"
540
- part["attribution_source"] = "ambiguous_female_person_and_part_margin"
 
 
 
 
 
 
 
541
  else:
542
  part["attribution"] = "ambiguous"
543
- part["attribution_source"] = "ambiguous_person_or_part_margin"
544
- elif part_margin_is_clear:
545
  # Preserve cropped-body detection when no person box exists, but do
546
- # not let a near-tie such as 0.250 vs 0.246 decide the whole image.
 
 
 
 
 
 
 
547
  part["attribution"] = part["raw_gender"]
548
- part["attribution_source"] = "contrastive_part_score_with_margin"
549
  else:
550
  part["attribution"] = "ambiguous"
551
  part["attribution_source"] = "insufficient_part_gender_margin"
@@ -604,6 +673,9 @@ def _run_grounding_dino(image: Image.Image) -> dict[str, Any]:
604
  "gender_margin": GENDER_MARGIN,
605
  "child_margin": CHILD_MARGIN,
606
  "part_gender_margin": PART_GENDER_MARGIN,
 
 
 
607
  "nms_iou_threshold": NMS_IOU_THRESHOLD,
608
  "match_iou_threshold": MATCH_IOU_THRESHOLD,
609
  "match_containment_threshold": MATCH_CONTAINMENT_THRESHOLD,
@@ -629,7 +701,7 @@ async def lifespan(_: FastAPI):
629
  app = FastAPI(
630
  title="AI Shield - Grounding DINO Base",
631
  description="Single-pass contrastive female-person and female body-part detector",
632
- version="1.4.0",
633
  lifespan=lifespan,
634
  )
635
 
 
27
  "male": ("man", "male person"),
28
  "child": ("young child", "baby", "toddler", "little kid"),
29
  }
30
+ PART_TYPES = (
31
+ "face",
32
+ "hand",
33
+ "chest",
34
+ "torso",
35
+ "abdomen",
36
+ "stomach",
37
+ "waist",
38
+ "belly",
39
+ "arm",
40
+ "body",
41
+ "leg",
42
+ "thigh",
43
+ "foot",
44
+ )
45
+ CENTRAL_PART_TYPES = frozenset(
46
+ {"chest", "torso", "abdomen", "stomach", "waist", "belly", "body"}
47
+ )
48
 
49
  PHRASE_SPECS: list[dict[str, str]] = []
50
  for phrase in PERSON_PHRASES["female"]:
 
88
  # A body part must also show a meaningful female/male separation before it can
89
  # decide the image on its own. This does not add another model forward pass.
90
  PART_GENDER_MARGIN = float(os.getenv("PART_GENDER_MARGIN", "0.05"))
91
+ # Central-body crops receive a slightly lower margin because they often lack a
92
+ # face or a complete person box.
93
+ CENTRAL_PART_GENDER_MARGIN = float(
94
+ os.getenv("CENTRAL_PART_GENDER_MARGIN", "0.04")
95
+ )
96
+ # "female hand" and "male hand" share a strong noun token, so their scores are
97
+ # often close. Retain a high female-hand score with a small positive separation.
98
+ STRONG_FEMALE_HAND_THRESHOLD = float(
99
+ os.getenv("STRONG_FEMALE_HAND_THRESHOLD", "0.40")
100
+ )
101
+ STRONG_FEMALE_HAND_MARGIN = float(os.getenv("STRONG_FEMALE_HAND_MARGIN", "0.01"))
102
  NMS_IOU_THRESHOLD = float(os.getenv("NMS_IOU_THRESHOLD", "0.75"))
103
  MATCH_IOU_THRESHOLD = float(os.getenv("MATCH_IOU_THRESHOLD", "0.25"))
104
  MATCH_CONTAINMENT_THRESHOLD = float(os.getenv("MATCH_CONTAINMENT_THRESHOLD", "0.80"))
 
116
  "GENDER_MARGIN": GENDER_MARGIN,
117
  "CHILD_MARGIN": CHILD_MARGIN,
118
  "PART_GENDER_MARGIN": PART_GENDER_MARGIN,
119
+ "CENTRAL_PART_GENDER_MARGIN": CENTRAL_PART_GENDER_MARGIN,
120
+ "STRONG_FEMALE_HAND_THRESHOLD": STRONG_FEMALE_HAND_THRESHOLD,
121
+ "STRONG_FEMALE_HAND_MARGIN": STRONG_FEMALE_HAND_MARGIN,
122
  "NMS_IOU_THRESHOLD": NMS_IOU_THRESHOLD,
123
  "MATCH_IOU_THRESHOLD": MATCH_IOU_THRESHOLD,
124
  "MATCH_CONTAINMENT_THRESHOLD": MATCH_CONTAINMENT_THRESHOLD,
 
548
 
549
  for part in body_parts:
550
  owner = _best_matching_person(part["box"], persons)
551
+ standard_part_margin_is_clear = part["score_margin"] >= PART_GENDER_MARGIN
552
+ central_part_margin_is_clear = (
553
+ part["part"] in CENTRAL_PART_TYPES
554
+ and part["score_margin"] >= CENTRAL_PART_GENDER_MARGIN
555
+ )
556
+ part_margin_is_clear = (
557
+ standard_part_margin_is_clear or central_part_margin_is_clear
558
+ )
559
+ strong_female_hand = (
560
+ part["part"] == "hand"
561
+ and part["raw_gender"] == "female"
562
+ and part["female_score"] >= STRONG_FEMALE_HAND_THRESHOLD
563
+ and part["score_margin"] >= STRONG_FEMALE_HAND_MARGIN
564
+ )
565
+ female_part_is_clear = (
566
+ part["raw_gender"] == "female"
567
+ and (part_margin_is_clear or strong_female_hand)
568
+ )
569
 
570
  if owner is not None:
571
  owner_hint = owner["winner"]
572
+ owner_hint_threshold = (
573
+ CHILD_THRESHOLD if owner_hint == "child" else PERSON_THRESHOLD
574
+ )
575
+ owner_hint_is_strong = owner["score"] >= owner_hint_threshold
576
  part["owner_gender"] = owner["gender"]
577
  part["owner_winner"] = owner_hint
578
  part["owner_confirmed"] = owner["confirmed"]
579
  part["owner_score"] = owner["score"]
580
+ part["owner_hint_strong"] = owner_hint_is_strong
581
 
582
  if owner["confirmed"]:
583
  part["attribution"] = owner["gender"]
584
  part["attribution_source"] = f"confirmed_{owner['gender']}_person_box"
585
+ elif owner_hint in {"male", "child"} and owner_hint_is_strong:
586
  # An ambiguous whole-person result still provides stronger
587
+ # context when its winning score reached the normal threshold
588
+ # but failed only the required score margin.
589
  part["attribution"] = owner_hint
590
  part["attribution_source"] = f"ambiguous_{owner_hint}_person_hint"
591
+ elif female_part_is_clear:
592
+ # A weak, unconfirmed person box must not override a strong
593
+ # cropped female-part detection.
594
  part["attribution"] = "female"
595
+ part["attribution_source"] = (
596
+ "female_hand_high_score"
597
+ if strong_female_hand and not part_margin_is_clear
598
+ else "female_part_gender_margin"
599
+ )
600
+ elif part_margin_is_clear:
601
+ part["attribution"] = part["raw_gender"]
602
+ part["attribution_source"] = "weak_person_part_gender_margin"
603
  else:
604
  part["attribution"] = "ambiguous"
605
+ part["attribution_source"] = "weak_person_and_part_margin"
606
+ elif female_part_is_clear:
607
  # Preserve cropped-body detection when no person box exists, but do
608
+ # not let an ordinary low-score near-tie decide the whole image.
609
+ part["attribution"] = "female"
610
+ part["attribution_source"] = (
611
+ "unowned_female_hand_high_score"
612
+ if strong_female_hand and not part_margin_is_clear
613
+ else "unowned_female_part_gender_margin"
614
+ )
615
+ elif part_margin_is_clear:
616
  part["attribution"] = part["raw_gender"]
617
+ part["attribution_source"] = "unowned_part_gender_margin"
618
  else:
619
  part["attribution"] = "ambiguous"
620
  part["attribution_source"] = "insufficient_part_gender_margin"
 
673
  "gender_margin": GENDER_MARGIN,
674
  "child_margin": CHILD_MARGIN,
675
  "part_gender_margin": PART_GENDER_MARGIN,
676
+ "central_part_gender_margin": CENTRAL_PART_GENDER_MARGIN,
677
+ "strong_female_hand_threshold": STRONG_FEMALE_HAND_THRESHOLD,
678
+ "strong_female_hand_margin": STRONG_FEMALE_HAND_MARGIN,
679
  "nms_iou_threshold": NMS_IOU_THRESHOLD,
680
  "match_iou_threshold": MATCH_IOU_THRESHOLD,
681
  "match_containment_threshold": MATCH_CONTAINMENT_THRESHOLD,
 
701
  app = FastAPI(
702
  title="AI Shield - Grounding DINO Base",
703
  description="Single-pass contrastive female-person and female body-part detector",
704
+ version="1.5.0",
705
  lifespan=lifespan,
706
  )
707