Update app.py
Browse files
app.py
CHANGED
|
@@ -38,9 +38,17 @@ PERSON_CLASS_SPECS: tuple[tuple[str, str, str, str], ...] = (
|
|
| 38 |
("female person", "person", "female", "person"),
|
| 39 |
("teenage girl", "person", "female", "person"),
|
| 40 |
("adolescent girl", "person", "female", "person"),
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
("adult man", "person", "male", "person"),
|
| 42 |
("man", "person", "male", "person"),
|
| 43 |
("male person", "person", "male", "person"),
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
("young child", "person", "child", "person"),
|
| 45 |
("baby", "person", "child", "person"),
|
| 46 |
("toddler", "person", "child", "person"),
|
|
@@ -84,6 +92,10 @@ def _build_class_specs() -> tuple[tuple[str, str, str, str], ...]:
|
|
| 84 |
(
|
| 85 |
("woman's bare abdomen", "part", "female", "abdomen"),
|
| 86 |
("man's bare abdomen", "part", "male", "abdomen"),
|
|
|
|
|
|
|
|
|
|
|
|
|
| 87 |
)
|
| 88 |
)
|
| 89 |
return tuple(specs)
|
|
@@ -146,9 +158,10 @@ OWNER_CONTAINMENT_THRESHOLD = float(
|
|
| 146 |
)
|
| 147 |
MIN_BOX_AREA_RATIO = float(os.getenv("MIN_BOX_AREA_RATIO", "0.0002"))
|
| 148 |
IMAGE_SIZE = int(os.getenv("IMAGE_SIZE", "768"))
|
| 149 |
-
TILE_IMAGE_SIZE = int(os.getenv("TILE_IMAGE_SIZE", "
|
| 150 |
TILE_OVERLAP = float(os.getenv("TILE_OVERLAP", "0.20"))
|
| 151 |
MAX_TILES = int(os.getenv("MAX_TILES", "4"))
|
|
|
|
| 152 |
ENABLE_TILE_PASS = os.getenv("ENABLE_TILE_PASS", "true").lower() not in {
|
| 153 |
"0",
|
| 154 |
"false",
|
|
@@ -196,6 +209,7 @@ if (
|
|
| 196 |
or TILE_IMAGE_SIZE <= 0
|
| 197 |
or not 0.0 <= TILE_OVERLAP < 0.5
|
| 198 |
or MAX_TILES <= 0
|
|
|
|
| 199 |
or MAX_DETECTIONS <= 0
|
| 200 |
):
|
| 201 |
raise RuntimeError("Invalid image-size, tiling, or detection-count configuration")
|
|
@@ -462,31 +476,11 @@ def _class_name(names: Any, class_id: int) -> str:
|
|
| 462 |
return ""
|
| 463 |
|
| 464 |
|
| 465 |
-
def
|
| 466 |
-
image: Image.Image
|
| 467 |
-
inference_size: int = IMAGE_SIZE,
|
| 468 |
) -> list[dict[str, Any]]:
|
| 469 |
-
|
| 470 |
-
image_array = np.asarray(image)
|
| 471 |
-
with torch.inference_mode():
|
| 472 |
-
results = model.predict(
|
| 473 |
-
source=image_array,
|
| 474 |
-
imgsz=inference_size,
|
| 475 |
-
conf=CANDIDATE_THRESHOLD,
|
| 476 |
-
iou=MODEL_IOU_THRESHOLD,
|
| 477 |
-
max_det=MAX_DETECTIONS,
|
| 478 |
-
device="cpu",
|
| 479 |
-
verbose=False,
|
| 480 |
-
agnostic_nms=False,
|
| 481 |
-
augment=False,
|
| 482 |
-
rect=True,
|
| 483 |
-
retina_masks=False,
|
| 484 |
-
predictor=MultiLabelDetectionPredictor,
|
| 485 |
-
)
|
| 486 |
-
|
| 487 |
-
if not results or results[0].boxes is None:
|
| 488 |
return []
|
| 489 |
-
result = results[0]
|
| 490 |
boxes = result.boxes
|
| 491 |
if len(boxes) == 0:
|
| 492 |
return []
|
|
@@ -524,6 +518,46 @@ def _raw_detections(
|
|
| 524 |
return detections
|
| 525 |
|
| 526 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 527 |
def _spatial_clusters(detections: list[dict[str, Any]]) -> list[list[dict[str, Any]]]:
|
| 528 |
clusters: list[list[dict[str, Any]]] = []
|
| 529 |
for detection in sorted(detections, key=lambda item: item["score"], reverse=True):
|
|
@@ -779,33 +813,39 @@ def _prioritize_tile_regions(
|
|
| 779 |
return sorted(regions, key=priority)
|
| 780 |
|
| 781 |
|
| 782 |
-
def
|
| 783 |
-
image: Image.Image,
|
| 784 |
-
) -> list[dict[str, Any]]:
|
| 785 |
-
|
| 786 |
-
(
|
| 787 |
-
|
| 788 |
-
|
| 789 |
-
|
| 790 |
-
|
| 791 |
-
|
| 792 |
-
|
| 793 |
-
detections = _raw_detections(tile, TILE_IMAGE_SIZE)
|
| 794 |
-
translated: list[dict[str, Any]] = []
|
| 795 |
-
for detection in detections:
|
| 796 |
-
box = detection["box"]
|
| 797 |
-
translated.append(
|
| 798 |
-
{
|
| 799 |
-
**detection,
|
| 800 |
-
"box": {
|
| 801 |
-
"x_min": round(box["x_min"] + region["x_min"], 2),
|
| 802 |
-
"y_min": round(box["y_min"] + region["y_min"], 2),
|
| 803 |
-
"x_max": round(box["x_max"] + region["x_min"], 2),
|
| 804 |
-
"y_max": round(box["y_max"] + region["y_min"], 2),
|
| 805 |
-
},
|
| 806 |
-
}
|
| 807 |
)
|
| 808 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 809 |
|
| 810 |
|
| 811 |
def _has_confirmed_female_evidence(raw: list[dict[str, Any]]) -> bool:
|
|
@@ -876,9 +916,11 @@ def _run_yoloe(image: Image.Image, include_debug: bool = False) -> dict[str, Any
|
|
| 876 |
analyzed_regions: list[dict[str, int]] = []
|
| 877 |
if tile_pass_used:
|
| 878 |
regions = _prioritize_tile_regions(_tile_regions(image), raw)
|
| 879 |
-
for
|
| 880 |
-
|
| 881 |
-
|
|
|
|
|
|
|
| 882 |
raw = first_pass_raw + tile_raw
|
| 883 |
if _has_confirmed_female_evidence(raw):
|
| 884 |
break
|
|
@@ -950,9 +992,13 @@ def _run_yoloe(image: Image.Image, include_debug: bool = False) -> dict[str, Any
|
|
| 950 |
},
|
| 951 |
"full_image_size": IMAGE_SIZE,
|
| 952 |
"tile_image_size": TILE_IMAGE_SIZE,
|
|
|
|
| 953 |
"tile_overlap": TILE_OVERLAP,
|
| 954 |
"tile_pass_used": tile_pass_used,
|
| 955 |
"tiles_analyzed": len(analyzed_regions),
|
|
|
|
|
|
|
|
|
|
| 956 |
"tile_regions": analyzed_regions,
|
| 957 |
"persons": people,
|
| 958 |
"body_parts": parts,
|
|
@@ -981,7 +1027,7 @@ async def lifespan(_: FastAPI):
|
|
| 981 |
app = FastAPI(
|
| 982 |
title="AI Shield - YOLOE-11-M Test",
|
| 983 |
description="Fixed-vocabulary female-person and female body-part detector",
|
| 984 |
-
version="0.
|
| 985 |
lifespan=lifespan,
|
| 986 |
)
|
| 987 |
|
|
|
|
| 38 |
("female person", "person", "female", "person"),
|
| 39 |
("teenage girl", "person", "female", "person"),
|
| 40 |
("adolescent girl", "person", "female", "person"),
|
| 41 |
+
("small woman", "person", "female", "person"),
|
| 42 |
+
("partially visible woman", "person", "female", "person"),
|
| 43 |
+
("woman lying down", "person", "female", "person"),
|
| 44 |
+
("woman taking a bath", "person", "female", "person"),
|
| 45 |
("adult man", "person", "male", "person"),
|
| 46 |
("man", "person", "male", "person"),
|
| 47 |
("male person", "person", "male", "person"),
|
| 48 |
+
("small man", "person", "male", "person"),
|
| 49 |
+
("partially visible man", "person", "male", "person"),
|
| 50 |
+
("man lying down", "person", "male", "person"),
|
| 51 |
+
("man taking a bath", "person", "male", "person"),
|
| 52 |
("young child", "person", "child", "person"),
|
| 53 |
("baby", "person", "child", "person"),
|
| 54 |
("toddler", "person", "child", "person"),
|
|
|
|
| 92 |
(
|
| 93 |
("woman's bare abdomen", "part", "female", "abdomen"),
|
| 94 |
("man's bare abdomen", "part", "male", "abdomen"),
|
| 95 |
+
("woman's upper body", "part", "female", "body"),
|
| 96 |
+
("man's upper body", "part", "male", "body"),
|
| 97 |
+
("woman's profile face", "part", "female", "face"),
|
| 98 |
+
("man's profile face", "part", "male", "face"),
|
| 99 |
)
|
| 100 |
)
|
| 101 |
return tuple(specs)
|
|
|
|
| 158 |
)
|
| 159 |
MIN_BOX_AREA_RATIO = float(os.getenv("MIN_BOX_AREA_RATIO", "0.0002"))
|
| 160 |
IMAGE_SIZE = int(os.getenv("IMAGE_SIZE", "768"))
|
| 161 |
+
TILE_IMAGE_SIZE = int(os.getenv("TILE_IMAGE_SIZE", "512"))
|
| 162 |
TILE_OVERLAP = float(os.getenv("TILE_OVERLAP", "0.20"))
|
| 163 |
MAX_TILES = int(os.getenv("MAX_TILES", "4"))
|
| 164 |
+
TILE_BATCH_SIZE = int(os.getenv("TILE_BATCH_SIZE", "2"))
|
| 165 |
ENABLE_TILE_PASS = os.getenv("ENABLE_TILE_PASS", "true").lower() not in {
|
| 166 |
"0",
|
| 167 |
"false",
|
|
|
|
| 209 |
or TILE_IMAGE_SIZE <= 0
|
| 210 |
or not 0.0 <= TILE_OVERLAP < 0.5
|
| 211 |
or MAX_TILES <= 0
|
| 212 |
+
or TILE_BATCH_SIZE <= 0
|
| 213 |
or MAX_DETECTIONS <= 0
|
| 214 |
):
|
| 215 |
raise RuntimeError("Invalid image-size, tiling, or detection-count configuration")
|
|
|
|
| 476 |
return ""
|
| 477 |
|
| 478 |
|
| 479 |
+
def _detections_from_result(
|
| 480 |
+
result: Any, image: Image.Image
|
|
|
|
| 481 |
) -> list[dict[str, Any]]:
|
| 482 |
+
if result.boxes is None:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 483 |
return []
|
|
|
|
| 484 |
boxes = result.boxes
|
| 485 |
if len(boxes) == 0:
|
| 486 |
return []
|
|
|
|
| 518 |
return detections
|
| 519 |
|
| 520 |
|
| 521 |
+
def _raw_detections_batch(
|
| 522 |
+
images: list[Image.Image], inference_size: int
|
| 523 |
+
) -> list[list[dict[str, Any]]]:
|
| 524 |
+
if not images:
|
| 525 |
+
return []
|
| 526 |
+
model: YOLOE = MODEL_DATA["model"]
|
| 527 |
+
sources = [np.asarray(image) for image in images]
|
| 528 |
+
with torch.inference_mode():
|
| 529 |
+
results = model.predict(
|
| 530 |
+
source=sources if len(sources) > 1 else sources[0],
|
| 531 |
+
imgsz=inference_size,
|
| 532 |
+
batch=len(sources),
|
| 533 |
+
conf=CANDIDATE_THRESHOLD,
|
| 534 |
+
iou=MODEL_IOU_THRESHOLD,
|
| 535 |
+
max_det=MAX_DETECTIONS,
|
| 536 |
+
device="cpu",
|
| 537 |
+
verbose=False,
|
| 538 |
+
agnostic_nms=False,
|
| 539 |
+
augment=False,
|
| 540 |
+
rect=True,
|
| 541 |
+
retina_masks=False,
|
| 542 |
+
predictor=MultiLabelDetectionPredictor,
|
| 543 |
+
)
|
| 544 |
+
if not results:
|
| 545 |
+
return [[] for _ in images]
|
| 546 |
+
parsed = [
|
| 547 |
+
_detections_from_result(result, image)
|
| 548 |
+
for result, image in zip(results, images)
|
| 549 |
+
]
|
| 550 |
+
if len(parsed) < len(images):
|
| 551 |
+
parsed.extend([] for _ in range(len(images) - len(parsed)))
|
| 552 |
+
return parsed
|
| 553 |
+
|
| 554 |
+
|
| 555 |
+
def _raw_detections(
|
| 556 |
+
image: Image.Image, inference_size: int = IMAGE_SIZE
|
| 557 |
+
) -> list[dict[str, Any]]:
|
| 558 |
+
return _raw_detections_batch([image], inference_size)[0]
|
| 559 |
+
|
| 560 |
+
|
| 561 |
def _spatial_clusters(detections: list[dict[str, Any]]) -> list[list[dict[str, Any]]]:
|
| 562 |
clusters: list[list[dict[str, Any]]] = []
|
| 563 |
for detection in sorted(detections, key=lambda item: item["score"], reverse=True):
|
|
|
|
| 813 |
return sorted(regions, key=priority)
|
| 814 |
|
| 815 |
|
| 816 |
+
def _tile_detections_batch(
|
| 817 |
+
image: Image.Image, regions: list[dict[str, int]]
|
| 818 |
+
) -> list[list[dict[str, Any]]]:
|
| 819 |
+
tiles = [
|
| 820 |
+
image.crop(
|
| 821 |
+
(
|
| 822 |
+
region["x_min"],
|
| 823 |
+
region["y_min"],
|
| 824 |
+
region["x_max"],
|
| 825 |
+
region["y_max"],
|
| 826 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 827 |
)
|
| 828 |
+
for region in regions
|
| 829 |
+
]
|
| 830 |
+
tile_results = _raw_detections_batch(tiles, TILE_IMAGE_SIZE)
|
| 831 |
+
translated_results: list[list[dict[str, Any]]] = []
|
| 832 |
+
for detections, region in zip(tile_results, regions):
|
| 833 |
+
translated: list[dict[str, Any]] = []
|
| 834 |
+
for detection in detections:
|
| 835 |
+
box = detection["box"]
|
| 836 |
+
translated.append(
|
| 837 |
+
{
|
| 838 |
+
**detection,
|
| 839 |
+
"box": {
|
| 840 |
+
"x_min": round(box["x_min"] + region["x_min"], 2),
|
| 841 |
+
"y_min": round(box["y_min"] + region["y_min"], 2),
|
| 842 |
+
"x_max": round(box["x_max"] + region["x_min"], 2),
|
| 843 |
+
"y_max": round(box["y_max"] + region["y_min"], 2),
|
| 844 |
+
},
|
| 845 |
+
}
|
| 846 |
+
)
|
| 847 |
+
translated_results.append(translated)
|
| 848 |
+
return translated_results
|
| 849 |
|
| 850 |
|
| 851 |
def _has_confirmed_female_evidence(raw: list[dict[str, Any]]) -> bool:
|
|
|
|
| 916 |
analyzed_regions: list[dict[str, int]] = []
|
| 917 |
if tile_pass_used:
|
| 918 |
regions = _prioritize_tile_regions(_tile_regions(image), raw)
|
| 919 |
+
for start in range(0, len(regions), TILE_BATCH_SIZE):
|
| 920 |
+
region_batch = regions[start : start + TILE_BATCH_SIZE]
|
| 921 |
+
analyzed_regions.extend(region_batch)
|
| 922 |
+
for detections in _tile_detections_batch(image, region_batch):
|
| 923 |
+
tile_raw.extend(detections)
|
| 924 |
raw = first_pass_raw + tile_raw
|
| 925 |
if _has_confirmed_female_evidence(raw):
|
| 926 |
break
|
|
|
|
| 992 |
},
|
| 993 |
"full_image_size": IMAGE_SIZE,
|
| 994 |
"tile_image_size": TILE_IMAGE_SIZE,
|
| 995 |
+
"tile_batch_size": TILE_BATCH_SIZE,
|
| 996 |
"tile_overlap": TILE_OVERLAP,
|
| 997 |
"tile_pass_used": tile_pass_used,
|
| 998 |
"tiles_analyzed": len(analyzed_regions),
|
| 999 |
+
"tile_batches_analyzed": math.ceil(
|
| 1000 |
+
len(analyzed_regions) / TILE_BATCH_SIZE
|
| 1001 |
+
),
|
| 1002 |
"tile_regions": analyzed_regions,
|
| 1003 |
"persons": people,
|
| 1004 |
"body_parts": parts,
|
|
|
|
| 1027 |
app = FastAPI(
|
| 1028 |
title="AI Shield - YOLOE-11-M Test",
|
| 1029 |
description="Fixed-vocabulary female-person and female body-part detector",
|
| 1030 |
+
version="0.4.0",
|
| 1031 |
lifespan=lifespan,
|
| 1032 |
)
|
| 1033 |
|