116 lines
3.9 KiB
Python
116 lines
3.9 KiB
Python
#!/opt/homebrew/bin/python3.11
|
|
"""
|
|
Face landmark QC: verify eyes/nose are within face bounding box.
|
|
Flags faces in DB where landmarks don't match the bbox.
|
|
Usage: python3 face_landmark_qc.py <file_uuid> [--threshold 0.5] [--fix]
|
|
"""
|
|
import sys, json, psycopg2, argparse
|
|
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("uuid")
|
|
parser.add_argument("--threshold", "-t", type=float, default=0.5,
|
|
help="Fraction of landmark points that must be inside bbox (default: 0.5)")
|
|
parser.add_argument("--fix", action="store_true", help="Update face_detections QC flag in DB")
|
|
args = parser.parse_args()
|
|
|
|
UUID = args.uuid
|
|
THRESHOLD = args.threshold
|
|
FACE_PATH = f"/Users/accusys/momentry/output_dev/{UUID}.face.json"
|
|
|
|
print(f"=== Face Landmark QC ===")
|
|
print(f"UUID: {UUID}")
|
|
print(f"Threshold: {THRESHOLD * 100:.0f}% points must be inside bbox")
|
|
|
|
# Load face.json
|
|
with open(FACE_PATH) as f:
|
|
data = json.load(f)
|
|
|
|
total_faces = 0
|
|
faces_with_lm = 0
|
|
good_faces = 0
|
|
bad_faces = 0
|
|
bad_frame_ids = set()
|
|
bad_face_details = []
|
|
|
|
# Build frame lookup for fast access
|
|
frame_map = {}
|
|
for frm in data['frames']:
|
|
frame_map[frm['frame']] = frm
|
|
|
|
for frame_num, frm in frame_map.items():
|
|
for fi, face in enumerate(frm.get('faces', [])):
|
|
total_faces += 1
|
|
lm = face.get('landmarks')
|
|
if not lm:
|
|
continue
|
|
faces_with_lm += 1
|
|
|
|
x, y, w, h = face['x'], face['y'], face['width'], face['height']
|
|
inside_pts = 0
|
|
total_pts = 0
|
|
eye_nose_inside = 0 # at least one point from each eye+nose inside
|
|
|
|
for lm_type in ['left_eye', 'right_eye', 'nose']:
|
|
points = lm.get(lm_type, [])
|
|
total_pts += len(points)
|
|
any_inside = False
|
|
for pt in points:
|
|
px, py = pt[0], pt[1]
|
|
if (x <= px <= x + w) and (y <= py <= y + h):
|
|
inside_pts += 1
|
|
any_inside = True
|
|
if any_inside:
|
|
eye_nose_inside += 1
|
|
|
|
ratio = inside_pts / max(1, total_pts)
|
|
|
|
if ratio >= THRESHOLD and eye_nose_inside >= 2:
|
|
good_faces += 1
|
|
else:
|
|
bad_faces += 1
|
|
bad_frame_ids.add(frame_num)
|
|
bad_face_details.append({
|
|
'frame': frame_num,
|
|
'face_idx': fi,
|
|
'bbox': [x, y, w, h],
|
|
'inside_pts': inside_pts,
|
|
'total_pts': total_pts,
|
|
'ratio': ratio,
|
|
'eye_nose_ok': eye_nose_inside,
|
|
})
|
|
|
|
print(f"\nTotal faces: {total_faces:,}")
|
|
print(f"Faces with landmarks: {faces_with_lm:,}")
|
|
print(f"✅ Good (≥{THRESHOLD*100:.0f}% inside + ≥2 features): {good_faces:,}")
|
|
print(f"❌ Bad: {bad_faces:,}")
|
|
print(f"Quality pass rate: {100 * good_faces / max(1, faces_with_lm):.1f}%")
|
|
|
|
print(f"\nBad faces in {len(bad_frame_ids)} unique frames")
|
|
|
|
# Show sample bad faces
|
|
print(f"\nSample bad faces:")
|
|
for bf in sorted(bad_face_details, key=lambda b: b['ratio'])[:5]:
|
|
print(f" frame={bf['frame']}, bbox={bf['bbox']}, {bf['inside_pts']}/{bf['total_pts']} inside ({bf['ratio']*100:.0f}%), eye/nose={bf['eye_nose_ok']}/3")
|
|
|
|
# Show sample good faces
|
|
print(f"\nSample good faces:")
|
|
good_details = []
|
|
for frame_num, frm in frame_map.items():
|
|
for face in frm.get('faces', []):
|
|
lm = face.get('landmarks')
|
|
if not lm:
|
|
continue
|
|
x, y, w, h = face['x'], face['y'], face['width'], face['height']
|
|
inside = sum(1 for pts in lm.values() for pt in pts
|
|
if (x <= pt[0] <= x + w) and (y <= pt[1] <= y + h))
|
|
total = sum(len(pts) for pts in lm.values())
|
|
if inside / max(1, total) >= THRESHOLD:
|
|
good_details.append((frame_num, x, y, w, h, inside, total))
|
|
if len(good_details) >= 5:
|
|
break
|
|
if len(good_details) >= 5:
|
|
break
|
|
|
|
for g in good_details:
|
|
print(f" frame={g[0]}, bbox=[{g[1]},{g[2]},{g[3]},{g[4]}], {g[5]}/{g[6]} inside ({100*g[5]/max(1,g[6]):.0f}%)")
|