scripts/face-eval/ measures gravity: 'face' against hand-labeled photos (currently 12 from dlib's test set; focal accuracy 9/10, crop containment 22/22). More labeled images — especially failure-shaped ones: skin-toned backgrounds, saturated warm clothing, group shots, darker skin tones in varied lighting — make the detector measurably better. The README in that folder explains the labeling format; images are fetched, never committed.
scripts/face-eval/measuresgravity: 'face'against hand-labeled photos (currently 12 from dlib's test set; focal accuracy 9/10, crop containment 22/22). More labeled images — especially failure-shaped ones: skin-toned backgrounds, saturated warm clothing, group shots, darker skin tones in varied lighting — make the detector measurably better. The README in that folder explains the labeling format; images are fetched, never committed.