Update app.py
Browse files
app.py
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
|
| 2 |
import io
|
| 3 |
import os
|
| 4 |
import socket
|
|
@@ -60,14 +60,17 @@ def _make_prompt() -> tuple[str, dict[str, tuple[int, int]]]:
|
|
| 60 |
|
| 61 |
DETECTION_PROMPT, PHRASE_CHAR_SPANS = _make_prompt()
|
| 62 |
|
| 63 |
-
#
|
| 64 |
-
#
|
| 65 |
CANDIDATE_THRESHOLD = float(os.getenv("CANDIDATE_THRESHOLD", "0.15"))
|
| 66 |
PERSON_THRESHOLD = float(os.getenv("PERSON_THRESHOLD", "0.25"))
|
| 67 |
PART_THRESHOLD = float(os.getenv("PART_THRESHOLD", "0.20"))
|
| 68 |
CHILD_THRESHOLD = float(os.getenv("CHILD_THRESHOLD", "0.25"))
|
| 69 |
GENDER_MARGIN = float(os.getenv("GENDER_MARGIN", "0.05"))
|
| 70 |
CHILD_MARGIN = float(os.getenv("CHILD_MARGIN", "0.05"))
|
|
|
|
|
|
|
|
|
|
| 71 |
NMS_IOU_THRESHOLD = float(os.getenv("NMS_IOU_THRESHOLD", "0.75"))
|
| 72 |
MATCH_IOU_THRESHOLD = float(os.getenv("MATCH_IOU_THRESHOLD", "0.25"))
|
| 73 |
MATCH_CONTAINMENT_THRESHOLD = float(os.getenv("MATCH_CONTAINMENT_THRESHOLD", "0.80"))
|
|
@@ -84,6 +87,7 @@ for threshold_name, threshold_value in {
|
|
| 84 |
"CHILD_THRESHOLD": CHILD_THRESHOLD,
|
| 85 |
"GENDER_MARGIN": GENDER_MARGIN,
|
| 86 |
"CHILD_MARGIN": CHILD_MARGIN,
|
|
|
|
| 87 |
"NMS_IOU_THRESHOLD": NMS_IOU_THRESHOLD,
|
| 88 |
"MATCH_IOU_THRESHOLD": MATCH_IOU_THRESHOLD,
|
| 89 |
"MATCH_CONTAINMENT_THRESHOLD": MATCH_CONTAINMENT_THRESHOLD,
|
|
@@ -304,6 +308,69 @@ def _boxes_match(first: dict[str, float], second: dict[str, float]) -> bool:
|
|
| 304 |
)
|
| 305 |
|
| 306 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 307 |
def _nms(detections: list[dict[str, Any]]) -> list[dict[str, Any]]:
|
| 308 |
kept: list[dict[str, Any]] = []
|
| 309 |
for detection in sorted(
|
|
@@ -396,6 +463,7 @@ def _run_grounding_dino(image: Image.Image) -> dict[str, Any]:
|
|
| 396 |
"label": winner if confirmed else "ambiguous person",
|
| 397 |
"kind": "person",
|
| 398 |
"gender": winner if confirmed else "ambiguous",
|
|
|
|
| 399 |
"score": round(winner_score, 4),
|
| 400 |
"female_score": round(scores["female"], 4),
|
| 401 |
"male_score": round(scores["male"], 4),
|
|
@@ -448,18 +516,39 @@ def _run_grounding_dino(image: Image.Image) -> dict[str, Any]:
|
|
| 448 |
body_parts.extend(_nms(part_candidates))
|
| 449 |
|
| 450 |
for part in body_parts:
|
| 451 |
-
|
| 452 |
-
|
| 453 |
-
|
| 454 |
-
|
| 455 |
-
|
| 456 |
-
part["
|
| 457 |
-
|
| 458 |
-
part["
|
| 459 |
-
part["
|
| 460 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 461 |
part["attribution"] = part["raw_gender"]
|
| 462 |
-
part["attribution_source"] = "
|
|
|
|
|
|
|
|
|
|
| 463 |
|
| 464 |
body_parts.sort(key=lambda item: item["score"], reverse=True)
|
| 465 |
female_parts = [item for item in body_parts if item["attribution"] == "female"]
|
|
@@ -514,6 +603,7 @@ def _run_grounding_dino(image: Image.Image) -> dict[str, Any]:
|
|
| 514 |
"child_threshold": CHILD_THRESHOLD,
|
| 515 |
"gender_margin": GENDER_MARGIN,
|
| 516 |
"child_margin": CHILD_MARGIN,
|
|
|
|
| 517 |
"nms_iou_threshold": NMS_IOU_THRESHOLD,
|
| 518 |
"match_iou_threshold": MATCH_IOU_THRESHOLD,
|
| 519 |
"match_containment_threshold": MATCH_CONTAINMENT_THRESHOLD,
|
|
@@ -539,7 +629,7 @@ async def lifespan(_: FastAPI):
|
|
| 539 |
app = FastAPI(
|
| 540 |
title="AI Shield - Grounding DINO Base",
|
| 541 |
description="Single-pass contrastive female-person and female body-part detector",
|
| 542 |
-
version="1.
|
| 543 |
lifespan=lifespan,
|
| 544 |
)
|
| 545 |
|
|
@@ -594,4 +684,4 @@ async def analyze(request: Request) -> dict[str, Any]:
|
|
| 594 |
@app.post("/analyze-file")
|
| 595 |
async def analyze_file(file: UploadFile = File(...)) -> dict[str, Any]:
|
| 596 |
image = await _read_uploaded_image(file)
|
| 597 |
-
return await _analyze_image(image)
|
|
|
|
| 1 |
+
|
| 2 |
import io
|
| 3 |
import os
|
| 4 |
import socket
|
|
|
|
| 60 |
|
| 61 |
DETECTION_PROMPT, PHRASE_CHAR_SPANS = _make_prompt()
|
| 62 |
|
| 63 |
+
# Low candidate/part thresholds retain weak cropped detections. The margins
|
| 64 |
+
# prevent close text-score ties from becoming confident classifications.
|
| 65 |
CANDIDATE_THRESHOLD = float(os.getenv("CANDIDATE_THRESHOLD", "0.15"))
|
| 66 |
PERSON_THRESHOLD = float(os.getenv("PERSON_THRESHOLD", "0.25"))
|
| 67 |
PART_THRESHOLD = float(os.getenv("PART_THRESHOLD", "0.20"))
|
| 68 |
CHILD_THRESHOLD = float(os.getenv("CHILD_THRESHOLD", "0.25"))
|
| 69 |
GENDER_MARGIN = float(os.getenv("GENDER_MARGIN", "0.05"))
|
| 70 |
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"))
|
|
|
|
| 87 |
"CHILD_THRESHOLD": CHILD_THRESHOLD,
|
| 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,
|
|
|
|
| 308 |
)
|
| 309 |
|
| 310 |
|
| 311 |
+
def _box_area(box: dict[str, float]) -> float:
|
| 312 |
+
return max(0.0, box["x_max"] - box["x_min"]) * max(
|
| 313 |
+
0.0, box["y_max"] - box["y_min"]
|
| 314 |
+
)
|
| 315 |
+
|
| 316 |
+
|
| 317 |
+
def _part_containment(
|
| 318 |
+
part_box: dict[str, float], person_box: dict[str, float]
|
| 319 |
+
) -> float:
|
| 320 |
+
intersection_width = max(
|
| 321 |
+
0.0,
|
| 322 |
+
min(part_box["x_max"], person_box["x_max"])
|
| 323 |
+
- max(part_box["x_min"], person_box["x_min"]),
|
| 324 |
+
)
|
| 325 |
+
intersection_height = max(
|
| 326 |
+
0.0,
|
| 327 |
+
min(part_box["y_max"], person_box["y_max"])
|
| 328 |
+
- max(part_box["y_min"], person_box["y_min"]),
|
| 329 |
+
)
|
| 330 |
+
part_area = _box_area(part_box)
|
| 331 |
+
return (
|
| 332 |
+
intersection_width * intersection_height / part_area
|
| 333 |
+
if part_area > 0.0
|
| 334 |
+
else 0.0
|
| 335 |
+
)
|
| 336 |
+
|
| 337 |
+
|
| 338 |
+
def _best_matching_person(
|
| 339 |
+
part_box: dict[str, float], persons: list[dict[str, Any]]
|
| 340 |
+
) -> dict[str, Any] | None:
|
| 341 |
+
matches = [person for person in persons if _boxes_match(part_box, person["box"])]
|
| 342 |
+
if not matches:
|
| 343 |
+
return None
|
| 344 |
+
|
| 345 |
+
# Keep matches that contain approximately the same amount of the part, then
|
| 346 |
+
# discard overly broad group boxes when a more specific person box exists.
|
| 347 |
+
# Within that local cluster, a confirmed person is more reliable than an
|
| 348 |
+
# ambiguous duplicate box around the same individual.
|
| 349 |
+
max_containment = max(
|
| 350 |
+
_part_containment(part_box, person["box"]) for person in matches
|
| 351 |
+
)
|
| 352 |
+
close_matches = [
|
| 353 |
+
person
|
| 354 |
+
for person in matches
|
| 355 |
+
if _part_containment(part_box, person["box"]) >= max_containment - 0.05
|
| 356 |
+
]
|
| 357 |
+
smallest_area = min(_box_area(person["box"]) for person in close_matches)
|
| 358 |
+
specific_matches = [
|
| 359 |
+
person
|
| 360 |
+
for person in close_matches
|
| 361 |
+
if _box_area(person["box"]) <= smallest_area * 4.0
|
| 362 |
+
]
|
| 363 |
+
return max(
|
| 364 |
+
specific_matches,
|
| 365 |
+
key=lambda person: (
|
| 366 |
+
person["confirmed"],
|
| 367 |
+
_part_containment(part_box, person["box"]),
|
| 368 |
+
-_box_area(person["box"]),
|
| 369 |
+
person["score"],
|
| 370 |
+
),
|
| 371 |
+
)
|
| 372 |
+
|
| 373 |
+
|
| 374 |
def _nms(detections: list[dict[str, Any]]) -> list[dict[str, Any]]:
|
| 375 |
kept: list[dict[str, Any]] = []
|
| 376 |
for detection in sorted(
|
|
|
|
| 463 |
"label": winner if confirmed else "ambiguous person",
|
| 464 |
"kind": "person",
|
| 465 |
"gender": winner if confirmed else "ambiguous",
|
| 466 |
+
"winner": winner,
|
| 467 |
"score": round(winner_score, 4),
|
| 468 |
"female_score": round(scores["female"], 4),
|
| 469 |
"male_score": round(scores["male"], 4),
|
|
|
|
| 516 |
body_parts.extend(_nms(part_candidates))
|
| 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"
|
| 552 |
|
| 553 |
body_parts.sort(key=lambda item: item["score"], reverse=True)
|
| 554 |
female_parts = [item for item in body_parts if item["attribution"] == "female"]
|
|
|
|
| 603 |
"child_threshold": CHILD_THRESHOLD,
|
| 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 |
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 |
|
|
|
|
| 684 |
@app.post("/analyze-file")
|
| 685 |
async def analyze_file(file: UploadFile = File(...)) -> dict[str, Any]:
|
| 686 |
image = await _read_uploaded_image(file)
|
| 687 |
+
return await _analyze_image(image)
|