Spaces:
Running
Running
File size: 9,309 Bytes
9f6a8e2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 | """Seat-vs-PIP-knuckle width validation over a folder of hand photos.
For each image: run the real pipeline up through canonical rotation, then
measure two bands on the chosen finger --
* SEAT : current anatomical band [PIP - seg , PIP] (median)
* KNUCKLE : PIP-centered band PIP +/- K_SEG*seg (median + p90)
-- map each to a ring size, and write
* an annotated overlay PNG so the seat vs knuckle lines can be eyeballed
* a row in results.csv
No ground truth is required: this is a visual + tabular comparison to confirm
the knuckle band detects the wider PIP joint on knuckle-dominant fingers.
Usage:
python script/validate_knuckle.py [input_dir] [--finger index] [--limit N]
"""
import sys
import csv
import argparse
from pathlib import Path
import numpy as np
import cv2
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from src.finger_segmentation import segment_hand, isolate_finger
from src.geometry import (
estimate_finger_axis, localize_ring_zone_from_landmarks,
calculate_angle_from_vertical, rotate_image_precise, rotate_axis_data,
transform_points_rotation,
)
from src.edge_refinement import refine_edges_sobel
from src.card_detection import compute_scale_factor
from src.ring_size import recommend_ring_size
from measure_finger import _sam_card_detect
# --- locked method parameters (per agreed recommendation) ---
K_SEG = 0.4 # knuckle band half-height in units of |DIP-PIP|
SLOPE, INTERCEPT = 0.809353, 0.251506 # src/calibration.json (mask/classic)
def _cal_mm(raw_px, px_per_cm):
raw_cm = raw_px / px_per_cm
return (SLOPE * raw_cm + INTERCEPT) * 10.0
def _size(cal_mm):
rs = recommend_ring_size(cal_mm / 10.0)
return rs["best_match"] if rs else None
def _band_widths_px(image, axis_data, zone, px_per_cm, landmarks, mask):
"""Return (median_px, p90_px, n) of per-row mask widths over a band."""
m = refine_edges_sobel(
image=image, axis_data=axis_data, zone_data=zone,
scale_px_per_cm=px_per_cm, finger_landmarks=landmarks,
finger_mask=mask, mask_mode="mask_only",
)
ed = m["edge_data"]
vr = ed["valid_rows"]
w = (ed["right_edges"][vr] - ed["left_edges"][vr])
if w.size == 0:
return None, None, 0
# same MAD trim the production median uses, so seat matches the real number
med = np.median(w)
mad = np.median(np.abs(w - med))
keep = np.abs(w - med) <= 3 * mad if mad > 0 else np.ones_like(w, bool)
w = w[keep]
return float(np.median(w)), float(np.percentile(w, 90)), int(w.size)
def _draw(img, landmarks, seat_c, seat_w, knk_c, knk_w, texts):
"""Draw both width lines + landmarks on a copy, crop tight, return it."""
vis = img.copy()
t = max(2, img.shape[1] // 500)
pts = [seat_c, knk_c] + [tuple(p) for p in landmarks]
def hline(c, half, color):
cx, cy = float(c[0]), float(c[1])
a = (int(cx - half), int(cy)); b = (int(cx + half), int(cy))
cv2.line(vis, a, b, color, t)
cv2.circle(vis, a, t + 1, color, -1)
cv2.circle(vis, b, t + 1, color, -1)
pts.extend([a, b])
hline(seat_c, seat_w / 2, (0, 200, 0)) # green = seat
hline(knk_c, knk_w / 2, (0, 80, 255)) # orange = knuckle
for name, p in zip(["MCP", "PIP", "DIP", "TIP"], landmarks):
cv2.circle(vis, (int(p[0]), int(p[1])), t + 2, (255, 255, 0), -1)
xs = [p[0] for p in pts]; ys = [p[1] for p in pts]
pad = int(img.shape[1] * 0.12)
x0 = max(0, int(min(xs)) - pad); x1 = min(img.shape[1], int(max(xs)) + pad)
y0 = max(0, int(min(ys)) - pad); y1 = min(img.shape[0], int(max(ys)) + pad)
crop = vis[y0:y1, x0:x1].copy()
fs = crop.shape[1] / 700.0
for i, txt in enumerate(texts):
y = int(30 * fs) + int(i * 34 * fs)
cv2.putText(crop, txt, (int(10 * fs), y), cv2.FONT_HERSHEY_SIMPLEX,
fs, (0, 0, 0), max(3, int(4 * fs)), cv2.LINE_AA)
cv2.putText(crop, txt, (int(10 * fs), y), cv2.FONT_HERSHEY_SIMPLEX,
fs, (255, 255, 255), max(1, int(2 * fs)), cv2.LINE_AA)
return crop
def process(path, finger, out_dir):
image = cv2.imread(str(path))
if image is None:
return {"name": path.stem, "fail_reason": "imread_failed"}
hand = segment_hand(image, finger=finger)
if hand is None:
return {"name": path.stem, "fail_reason": "hand_not_detected"}
img_can = hand.get("canonical_image", image)
raw_mask = hand.get("mask")
card = _sam_card_detect(img_can, hand, False, None)
if card is None:
return {"name": path.stem, "fail_reason": "card_not_detected"}
px_per_cm, _ = compute_scale_factor(card["corners"])
h, w = img_can.shape[:2]
fd = isolate_finger(hand, finger=finger, image_shape=(h, w))
if fd is None or fd.get("landmarks") is None:
return {"name": path.stem, "fail_reason": "finger_isolation_failed"}
axis = estimate_finger_axis(fd["landmarks"])
angle = calculate_angle_from_vertical(axis["direction"])
R = None
if abs(angle) >= 0.0:
img_can, R = rotate_image_precise(img_can, angle, (w / 2.0, h / 2.0))
axis = rotate_axis_data(axis, R)
lm = transform_points_rotation(fd["landmarks"], R)
raw_mask = cv2.warpAffine(raw_mask, R, (w, h), flags=cv2.INTER_NEAREST)
else:
lm = fd["landmarks"]
mcp, pip, dip, tip = lm
seg = float(np.linalg.norm(dip - pip))
direction = (pip - mcp) / np.linalg.norm(pip - mcp)
seat = localize_ring_zone_from_landmarks(lm, axis, zone_type="anatomical")
half = K_SEG * seg
knuckle = {
"start_point": (pip - direction * half).astype(np.float32),
"end_point": (pip + direction * half).astype(np.float32),
"center_point": pip.astype(np.float32),
"length": float(2 * half),
"localization_method": "pip_knuckle",
}
seat_med, _, seat_n = _band_widths_px(img_can, axis, seat, px_per_cm, lm, raw_mask)
knk_med, knk_p90, knk_n = _band_widths_px(img_can, axis, knuckle, px_per_cm, lm, raw_mask)
if seat_med is None or knk_med is None:
return {"name": path.stem, "fail_reason": "no_valid_rows"}
seat_cal = _cal_mm(seat_med, px_per_cm)
knk_cal = _cal_mm(knk_med, px_per_cm)
knk_p90_cal = _cal_mm(knk_p90, px_per_cm)
row = {
"name": path.stem,
"finger": finger,
"scale_px_per_cm": round(px_per_cm, 1),
"seat_cal_mm": round(seat_cal, 2),
"seat_size": _size(seat_cal),
"knuckle_med_cal_mm": round(knk_cal, 2),
"knuckle_size": _size(knk_cal),
"knuckle_p90_cal_mm": round(knk_p90_cal, 2),
"knuckle_p90_size": _size(knk_p90_cal),
"gap_mm": round(knk_cal - seat_cal, 2),
"size_up": int((_size(knk_cal) or 0) > (_size(seat_cal) or 0)),
"fail_reason": "",
}
texts = [
f"seat {seat_cal:.1f}mm size {row['seat_size']}",
f"knuckle {knk_cal:.1f}mm size {row['knuckle_size']} (gap {row['gap_mm']:+.1f}mm)",
]
crop = _draw(img_can, lm, seat["center_point"], seat_med,
pip, knk_med, texts)
cv2.imwrite(str(out_dir / f"{path.stem}__cmp.png"), crop)
return row
def main():
ap = argparse.ArgumentParser()
ap.add_argument("input_dir", nargs="?", default="input/kol_high_quality")
ap.add_argument("--finger", default="index")
ap.add_argument("--limit", type=int, default=0)
args = ap.parse_args()
in_dir = Path(args.input_dir)
out_dir = Path("output/knuckle_validation")
out_dir.mkdir(parents=True, exist_ok=True)
exts = {".jpg", ".jpeg", ".png"}
images = sorted(p for p in in_dir.iterdir() if p.suffix.lower() in exts)
if args.limit:
images = images[:args.limit]
fields = ["name", "finger", "scale_px_per_cm", "seat_cal_mm", "seat_size",
"knuckle_med_cal_mm", "knuckle_size", "knuckle_p90_cal_mm",
"knuckle_p90_size", "gap_mm", "size_up", "fail_reason"]
rows = []
for i, p in enumerate(images, 1):
try:
r = process(p, args.finger, out_dir)
except Exception as e:
r = {"name": p.stem, "fail_reason": f"error:{type(e).__name__}:{e}"}
rows.append({k: r.get(k, "") for k in fields})
print(f"[{i}/{len(images)}] {p.stem}: "
+ (r["fail_reason"] if r.get("fail_reason")
else f"seat={r['seat_size']} knuckle={r['knuckle_size']} gap={r['gap_mm']:+.1f}mm"),
flush=True)
csv_path = out_dir / "results.csv"
with open(csv_path, "w", newline="") as f:
wcsv = csv.DictWriter(f, fieldnames=fields)
wcsv.writeheader()
wcsv.writerows(rows)
ok = [r for r in rows if not r["fail_reason"]]
up = [r for r in ok if r["size_up"]]
print("\n==== summary ====")
print(f"processed {len(rows)}, ok {len(ok)}, failed {len(rows) - len(ok)}")
if ok:
gaps = [float(r["gap_mm"]) for r in ok]
print(f"knuckle wider in {sum(g > 0 for g in gaps)}/{len(ok)} images; "
f"mean gap {np.mean(gaps):+.2f}mm, max {max(gaps):+.2f}mm")
print(f"knuckle bumps the recommended size up in {len(up)}/{len(ok)} images")
print(f"overlays + results.csv -> {out_dir}")
if __name__ == "__main__":
main()
|