File size: 37,454 Bytes
58dd7d3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 | """
COMPLETE PATCHED preprocessing_pipeline.py
ALL FIXES INTEGRATED:
- Anisotropic foreshortening correction in compute_roi_warp()
- Updated step1_normalize() to pass head_pose
- Fixed bounds check in extract_edge_points()
STEP 2 - Region-Aware Illumination Normalization
=================================================
Novel Eye Tracking Preprocessing Pipeline | InsightUX
Run AFTER step1_geometric_normalization.py.
This file imports Step 1 and adds Step 2 on top.
"""
import cv2
import numpy as np
import mediapipe as mp
import argparse
import sys
from dataclasses import dataclass
from typing import Tuple, Optional
# =============================================================================
# MediaPipe + Iris Radius Helpers
# =============================================================================
def create_face_mesh(static_image_mode: bool = False,
max_num_faces: int = 1,
refine_landmarks: bool = True,
min_detection_confidence: float = 0.5,
min_tracking_confidence: float = 0.5):
"""
Creates and returns a MediaPipe FaceMesh instance.
refine_landmarks=True is required for iris landmarks (indices 468-477).
"""
try:
from mediapipe.python.solutions.face_mesh import FaceMesh
except ImportError as exc:
py = sys.version.split()[0]
mp_ver = getattr(mp, "__version__", "unknown")
raise RuntimeError(
f"mediapipe {mp_ver} on Python {py} does not include the legacy "
"solutions.face_mesh API. Use Python 3.11 with mediapipe==0.10.18 "
"and reinstall: pip install mediapipe==0.10.18"
) from exc
return FaceMesh(
static_image_mode=static_image_mode,
max_num_faces=max_num_faces,
refine_landmarks=refine_landmarks,
min_detection_confidence=min_detection_confidence,
min_tracking_confidence=min_tracking_confidence,
)
def compute_iris_radius(landmarks, iris_indices: list, frame_shape: tuple) -> float:
"""
Computes iris radius in frame pixels from MediaPipe iris landmarks.
MediaPipe iris layout: index 0 = center, 1-4 = cardinal rim points.
Radius = mean distance from center to the 4 rim points.
ALWAYS use this instead of hardcoding iris_radius_frame = 15.0.
"""
h, w = frame_shape[:2]
pts = np.array(
[(landmarks[i].x * w, landmarks[i].y * h) for i in iris_indices],
dtype=np.float32
)
center = pts[0]
radius = float(np.mean(np.linalg.norm(pts[1:] - center, axis=1)))
return max(radius, 1.0)
# -----------------------------------------------------------------------------
# 3D Face Model + Landmark Indices
# -----------------------------------------------------------------------------
# Generic 6-point 3D face model in mm (origin = nose tip)
#
# FIX (head-pose convention bug): the original model used a Y-up
# convention (chin negative, eyes positive), which conflicts with
# OpenCV's camera convention (Y-down, Z-forward). That mismatch made
# solvePnP report a baked-in ~180deg pitch offset for a straight,
# frontal face (observed as head_pose.pitch landing near -177deg
# instead of a sane small value). Y and Z signs are flipped here so
# the model matches OpenCV's convention: a frontal face now yields
# head_pose.pitch/yaw/roll near 0, as it should.
FACE_3D_MODEL = np.array([
[ 0.0, 0.0, 0.0], # nose tip -> landmark 1
[ 0.0, 63.6, 12.5], # chin -> landmark 152
[-43.3, -32.7, 26.0], # left eye outer -> landmark 33
[ 43.3, -32.7, 26.0], # right eye outer -> landmark 263
[-28.9, 28.9, 24.1], # left mouth -> landmark 61
[ 28.9, 28.9, 24.1], # right mouth -> landmark 291
], dtype=np.float64)
FACE_2D_INDICES = [1, 152, 33, 263, 61, 291]
LEFT_EYE_INDICES = [33, 7, 163, 144, 145, 153, 154, 155,
133, 173, 157, 158, 159, 160, 161, 246]
RIGHT_EYE_INDICES = [362, 382, 381, 380, 374, 373, 390, 249,
263, 466, 388, 387, 386, 385, 384, 398]
LEFT_IRIS_INDICES = [468, 469, 470, 471, 472] # [center, top, right, bottom, left]
RIGHT_IRIS_INDICES = [473, 474, 475, 476, 477]
LEFT_EAR_INDICES = [33, 160, 158, 133, 153, 144]
RIGHT_EAR_INDICES = [362, 385, 387, 263, 373, 380]
PATCH_W, PATCH_H = 60, 36
EYE_PAD_FACTOR = 0.45
# -----------------------------------------------------------------------------
# Data Classes
# -----------------------------------------------------------------------------
@dataclass
class HeadPose:
pitch: float
yaw: float
roll: float
rvec: np.ndarray
tvec: np.ndarray
reproj_error: float
@dataclass
class NormalizedEyePatch:
raw_crop: np.ndarray
norm_crop: np.ndarray
diff_map: np.ndarray
bbox_raw: Tuple[int,int,int,int]
warp_M: np.ndarray
iris_center: Tuple[float,float]
ear: float
is_open: bool
norm_quality: float
# -----------------------------------------------------------------------------
# Camera Intrinsics
# -----------------------------------------------------------------------------
def estimate_camera_matrix(frame_shape: tuple) -> np.ndarray:
h, w = frame_shape[:2]
f = float(w)
cx = w / 2.0
cy = h / 2.0
return np.array([
[f, 0., cx],
[0., f, cy],
[0., 0., 1.],
], dtype=np.float64)
# -----------------------------------------------------------------------------
# Head Pose Estimation
# -----------------------------------------------------------------------------
def estimate_head_pose(
landmarks,
frame_shape: tuple,
camera_matrix: np.ndarray,
) -> Optional[HeadPose]:
h, w = frame_shape[:2]
dist = np.zeros((4, 1), dtype=np.float64)
pts2d = np.array(
[(landmarks[i].x * w, landmarks[i].y * h) for i in FACE_2D_INDICES],
dtype=np.float64
)
ok, rvec, tvec = cv2.solvePnP(
FACE_3D_MODEL, pts2d, camera_matrix, dist,
flags=cv2.SOLVEPNP_EPNP
)
if not ok:
return None
rvec, tvec = cv2.solvePnPRefineLM(
FACE_3D_MODEL, pts2d, camera_matrix, dist, rvec, tvec
)
proj, _ = cv2.projectPoints(FACE_3D_MODEL, rvec, tvec, camera_matrix, dist)
reproj = float(np.mean(np.linalg.norm(proj.reshape(-1,2) - pts2d, axis=1)))
R, _ = cv2.Rodrigues(rvec)
pitch, yaw, roll = _R_to_euler(R)
return HeadPose(pitch=pitch, yaw=yaw, roll=roll,
rvec=rvec, tvec=tvec, reproj_error=reproj)
def _R_to_euler(R: np.ndarray) -> Tuple[float, float, float]:
"""Rotation matrix -> (pitch, yaw, roll) degrees."""
sy = np.sqrt(R[0,0]**2 + R[1,0]**2)
if sy > 1e-6:
pitch = np.degrees(np.arctan2( R[2,1], R[2,2]))
yaw = np.degrees(np.arctan2(-R[2,0], sy))
roll = np.degrees(np.arctan2( R[1,0], R[0,0]))
else:
pitch = np.degrees(np.arctan2(-R[1,2], R[1,1]))
yaw = np.degrees(np.arctan2(-R[2,0], sy))
roll = 0.0
return pitch, yaw, roll
# -----------------------------------------------------------------------------
# ROI Extraction
# -----------------------------------------------------------------------------
def extract_roi(
frame: np.ndarray,
landmarks,
eye_indices: list,
pad: float = EYE_PAD_FACTOR,
) -> Tuple[np.ndarray, Tuple[int,int,int,int]]:
h, w = frame.shape[:2]
pts = np.array(
[(landmarks[i].x * w, landmarks[i].y * h) for i in eye_indices],
dtype=np.float32
)
xmin, ymin = pts.min(axis=0)
xmax, ymax = pts.max(axis=0)
ew, eh = xmax - xmin, ymax - ymin
x1 = max(0, int(xmin - ew * pad))
y1 = max(0, int(ymin - eh * pad))
x2 = min(w, int(xmax + ew * pad))
y2 = min(h, int(ymax + eh * pad))
if x2 <= x1 or y2 <= y1:
blank = np.zeros((PATCH_H, PATCH_W), dtype=np.uint8)
return blank, (0, 0, PATCH_W, PATCH_H)
roi = frame[y1:y2, x1:x2]
if roi.ndim == 3:
roi = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)
return roi.copy(), (x1, y1, x2, y2)
# =============================================================================
# 2D Normalization Warp (PATCHED WITH FORESHORTENING CORRECTION)
# =============================================================================
def compute_roi_warp(
landmarks,
frame_shape: tuple,
bbox: Tuple[int,int,int,int],
anchor_indices,
roll_deg: float,
head_pose: 'HeadPose' = None,
) -> Tuple[np.ndarray, Tuple[float,float]]:
"""
Compute 2D affine warp for eye ROI normalization, with optional
anisotropic foreshortening correction.
PATCHED: Now includes foreshortening correction via anisotropic pre-scaling.
Args:
head_pose: optional HeadPose object. If provided, applies anisotropic
pre-scaling to approximately undo the eye shape distortion
caused by head pitch/yaw. If None, uses the old (roll-only)
normalization.
Returns:
(warp_M, anchor_local): affine warp matrix and anchor point in ROI coords
The foreshortening correction works by:
- A head pitched down (looking at screen) makes the eye appear taller
in the image (Y foreshortened) -> pre-scale X by 1/cos(pitch)
- A head turned left (yaw negative) makes the eye appear narrower
in the image (X foreshortened) -> pre-scale Y by 1/cos(yaw)
Both corrections are applied as pre-scaling before the existing similarity
(rotate + translate + uniform scale) warp, so they don't interfere with
roll correction.
"""
h, w = frame_shape[:2]
x1, y1, x2, y2 = bbox
roi_w = max(x2 - x1, 1)
roi_h = max(y2 - y1, 1)
anchor_pts = np.array(
[(landmarks[i].x * w, landmarks[i].y * h) for i in anchor_indices],
dtype=np.float32
)
anchor_global = anchor_pts.mean(axis=0)
anchor_local = (
float(anchor_global[0] - x1),
float(anchor_global[1] - y1),
)
# ===== ANISOTROPIC PRE-SCALING FOR FORESHORTENING CORRECTION (PATCHED) =====
prescale_x = 1.0
prescale_y = 1.0
if head_pose is not None:
# Pitch correction: pitched-down face makes eye appear taller (Y compressed)
# -> pre-scale X to compensate
pitch_rad = np.radians(np.clip(head_pose.pitch, -35.0, 35.0))
prescale_x *= 1.0 / (np.cos(pitch_rad) + 1e-6)
# Yaw correction: turned-away face makes eye appear narrower (X compressed)
# -> pre-scale Y to compensate
yaw_rad = np.radians(np.clip(head_pose.yaw, -35.0, 35.0))
prescale_y *= 1.0 / (np.cos(yaw_rad) + 1e-6)
# Clip to prevent extreme scaling (edge case: head >45° off-axis)
prescale_x = np.clip(prescale_x, 0.8, 1.5)
prescale_y = np.clip(prescale_y, 0.8, 1.5)
# =========================================================================
# Effective ROI dimensions after pre-scaling
roi_w_eff = roi_w * prescale_x
roi_h_eff = roi_h * prescale_y
scale = min(PATCH_W / roi_w_eff, PATCH_H / roi_h_eff)
a = np.radians(-roll_deg)
cos_a, sin_a = np.cos(a), np.sin(a)
cx, cy = anchor_local
# Apply pre-scaling to anchor point before rotation/translation
cx_prescaled = cx * prescale_x
cy_prescaled = cy * prescale_y
tx = PATCH_W / 2.0 + scale * (-cos_a * cx_prescaled + sin_a * cy_prescaled)
ty = PATCH_H / 2.0 + scale * (-sin_a * cx_prescaled - cos_a * cy_prescaled)
warp_M = np.array([
[scale * cos_a, -scale * sin_a, tx],
[scale * sin_a, scale * cos_a, ty],
], dtype=np.float64)
return warp_M, anchor_local
def apply_warp_to_roi(
gray_roi: np.ndarray,
warp_M: np.ndarray,
) -> np.ndarray:
return cv2.warpAffine(
gray_roi, warp_M, (PATCH_W, PATCH_H),
flags=cv2.INTER_LANCZOS4,
borderMode=cv2.BORDER_REPLICATE
)
def raw_crop_resized(gray_roi: np.ndarray) -> np.ndarray:
if gray_roi.size == 0:
return np.zeros((PATCH_H, PATCH_W), dtype=np.uint8)
return cv2.resize(gray_roi, (PATCH_W, PATCH_H), interpolation=cv2.INTER_LANCZOS4)
# =============================================================================
# Iris Center in Normalized Patch Coords
# =============================================================================
def iris_center_in_patch(
iris_local: Tuple[float,float],
warp_M: np.ndarray,
) -> Tuple[float,float]:
ix, iy = iris_local
px = warp_M[0,0]*ix + warp_M[0,1]*iy + warp_M[0,2]
py = warp_M[1,0]*ix + warp_M[1,1]*iy + warp_M[1,2]
return float(px), float(py)
# =============================================================================
# Normalization Quality Metric
# =============================================================================
def compute_norm_quality(
raw_crop: np.ndarray,
norm_crop: np.ndarray,
) -> float:
def entropy(patch: np.ndarray) -> float:
gx = cv2.Sobel(patch, cv2.CV_64F, 1, 0, ksize=3)
gy = cv2.Sobel(patch, cv2.CV_64F, 0, 1, ksize=3)
angles = np.arctan2(gy, gx)
hist, _ = np.histogram(angles.ravel(), bins=36, range=(-np.pi, np.pi))
hist = hist.astype(np.float64) + 1e-9
hist /= hist.sum()
return float(-np.sum(hist * np.log(hist)))
H_raw = entropy(raw_crop)
H_norm = entropy(norm_crop)
if H_raw < 1e-9:
return 0.0
return float(np.clip((H_raw - H_norm) / H_raw, 0.0, 1.0))
# =============================================================================
# EAR
# =============================================================================
def compute_ear(landmarks, ear_indices: list, frame_shape: tuple) -> float:
h, w = frame_shape[:2]
pts = np.array(
[(landmarks[i].x * w, landmarks[i].y * h) for i in ear_indices],
dtype=np.float32
)
A = np.linalg.norm(pts[1] - pts[5])
B = np.linalg.norm(pts[2] - pts[4])
C = np.linalg.norm(pts[0] - pts[3])
return float((A + B) / (2.0 * C)) if C > 1e-6 else 0.0
# =============================================================================
# Master Step 1 Function (PATCHED TO PASS head_pose)
# =============================================================================
def step1_normalize(
frame: np.ndarray,
landmarks,
head_pose: HeadPose,
eye_indices: list,
ear_indices: list,
iris_indices: list,
ear_threshold: float = 0.20,
) -> NormalizedEyePatch:
"""
PATCHED: Now passes full head_pose to compute_roi_warp for foreshortening correction.
"""
ear = compute_ear(landmarks, ear_indices, frame.shape)
is_open = ear >= ear_threshold
gray_roi, bbox = extract_roi(frame, landmarks, eye_indices)
# ANCHOR FIX: warp now centers on the eye-corner midpoint, not the
# iris. ear_indices[0] and ear_indices[3] are the outer and inner
# canthus (eye corner) landmarks - see compute_ear, where they're
# used as the EAR denominator (eye width).
#
# PATCHED: Also pass full head_pose to compute_roi_warp for foreshortening
# correction (was only passing roll before)
canthus_indices = (ear_indices[0], ear_indices[3])
warp_M, _anchor_local = compute_roi_warp(
landmarks, frame.shape, bbox, canthus_indices, head_pose.roll,
head_pose=head_pose # PATCHED: pass full head pose for foreshortening correction
)
norm_crop = apply_warp_to_roi(gray_roi, warp_M)
raw_crop = raw_crop_resized(gray_roi)
diff = cv2.absdiff(raw_crop, norm_crop)
diff_vis = cv2.applyColorMap(
cv2.convertScaleAbs(diff, alpha=4.0), cv2.COLORMAP_INFERNO
)
# Iris position in ROI-local coords, projected through the SAME warp.
# Still needed for Step 2's region-aware CLAHE (iris vs sclera mask).
# It now lands at a different spot in the patch depending on gaze
# direction, instead of always sitting dead-center.
h, w = frame.shape[:2]
iris_pts = np.array(
[(landmarks[i].x * w, landmarks[i].y * h) for i in iris_indices],
dtype=np.float32
)
iris_global = iris_pts[0]
x1, y1, _, _ = bbox
iris_local = (float(iris_global[0] - x1), float(iris_global[1] - y1))
iris_patch = iris_center_in_patch(iris_local, warp_M)
quality = compute_norm_quality(raw_crop, norm_crop) if is_open else 0.0
return NormalizedEyePatch(
raw_crop=raw_crop,
norm_crop=norm_crop,
diff_map=diff_vis,
bbox_raw=bbox,
warp_M=warp_M,
iris_center=iris_patch,
ear=ear,
is_open=is_open,
norm_quality=quality,
)
# =============================================================================
# Visualization helpers (used only by the standalone run_step1/2/3 viewers)
# =============================================================================
def draw_pose_axes(
frame: np.ndarray,
head_pose: HeadPose,
camera_matrix: np.ndarray,
nose_lm,
frame_shape: tuple,
length: float = 50.0,
) -> None:
h, w = frame_shape[:2]
dist = np.zeros((4,1))
ax3d = np.float32([[length,0,0],[0,length,0],[0,0,length]])
pts,_ = cv2.projectPoints(ax3d, head_pose.rvec, head_pose.tvec,
camera_matrix, dist)
pts = pts.reshape(-1,2).astype(int)
orig = (int(nose_lm.x * w), int(nose_lm.y * h))
cv2.arrowedLine(frame, orig, tuple(pts[0]), (0,0,220), 2, tipLength=0.2)
cv2.arrowedLine(frame, orig, tuple(pts[1]), (0,220,0), 2, tipLength=0.2)
cv2.arrowedLine(frame, orig, tuple(pts[2]), (220,100,0), 2, tipLength=0.2)
def draw_main_view(
frame: np.ndarray,
left: NormalizedEyePatch,
right: NormalizedEyePatch,
head_pose: HeadPose,
camera_matrix: np.ndarray,
landmarks,
frame_shape: tuple,
) -> np.ndarray:
out = frame.copy()
for patch, lbl in [(left,"L"), (right,"R")]:
x1,y1,x2,y2 = patch.bbox_raw
col = (0,255,0) if patch.is_open else (0,0,255)
cv2.rectangle(out,(x1,y1),(x2,y2),col,1)
cv2.putText(out, f"{lbl} EAR:{patch.ear:.2f} Q:{patch.norm_quality:.3f}",
(x1, max(0,y1-6)), cv2.FONT_HERSHEY_SIMPLEX, 0.38, col, 1)
draw_pose_axes(out, head_pose, camera_matrix, landmarks[1], frame_shape)
lines = [
f"Pitch:{head_pose.pitch:+.1f} Yaw:{head_pose.yaw:+.1f} Roll:{head_pose.roll:+.1f}",
f"Reproj err: {head_pose.reproj_error:.1f}px",
]
for i, ln in enumerate(lines):
cv2.putText(out, ln, (10, 20+i*18),
cv2.FONT_HERSHEY_SIMPLEX, 0.45, (200,230,255), 1)
return out
# =============================================================================
# CLAHE PARAMETERS (tuned per region)
# =============================================================================
CLAHE_IRIS = cv2.createCLAHE(clipLimit=1.5, tileGridSize=(2, 2))
CLAHE_SCLERA = cv2.createCLAHE(clipLimit=3.5, tileGridSize=(4, 4))
CLAHE_GLOBAL = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(4, 4))
# =============================================================================
# Step 2 Data Class
# =============================================================================
@dataclass
class IlluminationResult:
step1: NormalizedEyePatch
iris_mask: np.ndarray
sclera_mask: np.ndarray
iris_clahe: np.ndarray
sclera_clahe: np.ndarray
blended: np.ndarray
global_clahe: np.ndarray
diff_vs_global: np.ndarray
diff_vs_raw: np.ndarray
itv: float
photo_quality: float
is_usable: bool
ITV_THRESHOLD = 0.25
# =============================================================================
# Stage 2A - Adaptive Glint Removal
# =============================================================================
def adaptive_glint_threshold(patch: np.ndarray) -> int:
high_vals = patch[patch > np.percentile(patch, 95)]
if len(high_vals) < 10:
return 240
bright_norm = ((high_vals - high_vals.min()) /
(high_vals.ptp() + 1e-6) * 255).astype(np.uint8)
thresh, _ = cv2.threshold(
bright_norm, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU
)
adapted = int(high_vals.min() + thresh / 255.0 * high_vals.ptp())
return max(200, min(adapted, 254))
def remove_glints_adaptive(patch: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
thresh = adaptive_glint_threshold(patch)
_, glint_mask = cv2.threshold(patch, thresh, 255, cv2.THRESH_BINARY)
if cv2.countNonZero(glint_mask) == 0:
return patch.copy(), glint_mask
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
glint_mask = cv2.dilate(glint_mask, kernel, iterations=1)
cleaned = cv2.inpaint(patch, glint_mask, inpaintRadius=3,
flags=cv2.INPAINT_TELEA)
return cleaned, glint_mask
# =============================================================================
# Stage 2B - Iris Mask Construction
# =============================================================================
def build_iris_mask(
iris_center_patch: Tuple[float, float],
iris_radius_patch: float,
patch_shape: Tuple[int, int],
feather_width: float = 0.35,
) -> np.ndarray:
H, W = patch_shape
cx, cy = iris_center_patch
r_iris = max(iris_radius_patch, 1.0)
ys, xs = np.mgrid[0:H, 0:W]
dist = np.sqrt((xs - cx)**2 + (ys - cy)**2).astype(np.float32)
r_inner = r_iris * (1.0 - feather_width)
r_outer = r_iris
mask = np.zeros((H, W), dtype=np.float32)
mask[dist <= r_inner] = 1.0
feather_zone = (dist > r_inner) & (dist <= r_outer)
if feather_zone.any():
t = (dist[feather_zone] - r_inner) / (r_outer - r_inner + 1e-6)
mask[feather_zone] = np.cos(0.5 * np.pi * t) ** 2
return mask
def estimate_iris_radius_in_patch(
iris_radius_frame: float,
warp_M: np.ndarray,
) -> float:
scale = float(np.linalg.norm(warp_M[0, :2]))
return iris_radius_frame * scale
# =============================================================================
# Stage 2C - Region-Aware CLAHE
# =============================================================================
def region_aware_clahe(
patch: np.ndarray,
iris_mask: np.ndarray,
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
iris_enhanced = CLAHE_IRIS.apply(patch)
sclera_enhanced = CLAHE_SCLERA.apply(patch)
alpha = iris_mask[:, :, np.newaxis] if patch.ndim == 3 else iris_mask
blended = (
alpha * iris_enhanced.astype(np.float32) +
(1 - alpha) * sclera_enhanced.astype(np.float32)
).clip(0, 255).astype(np.uint8)
blended = cv2.bilateralFilter(blended, d=5, sigmaColor=20, sigmaSpace=3)
return iris_enhanced, sclera_enhanced, blended
# =============================================================================
# Stage 2D - Photometric Quality Metrics
# =============================================================================
def iris_texture_visibility(
blended: np.ndarray,
iris_mask: np.ndarray,
) -> float:
log_response = cv2.Laplacian(blended, cv2.CV_64F, ksize=3)
mask_sum = iris_mask.sum() + 1e-6
weighted_mean = float((log_response * iris_mask).sum() / mask_sum)
weighted_var = float(
((log_response - weighted_mean)**2 * iris_mask).sum() / mask_sum
)
itv = 1.0 / (1.0 + np.exp(-(weighted_var - 80.0) / 30.0))
return float(np.clip(itv, 0.0, 1.0))
def photometric_quality(
blended: np.ndarray,
iris_mask: np.ndarray,
glint_mask: np.ndarray,
) -> float:
itv = iris_texture_visibility(blended, iris_mask)
iris_area = float(iris_mask.sum()) + 1e-6
glint_in_iris = float(((glint_mask > 0).astype(np.float32) * iris_mask).sum())
glint_penalty = np.clip(glint_in_iris / iris_area, 0.0, 1.0)
mean_bright = float(blended.mean())
exposure_score = 1.0 - abs(mean_bright - 120.0) / 120.0
exposure_score = float(np.clip(exposure_score, 0.0, 1.0))
quality = 0.6 * itv + 0.2 * (1 - glint_penalty) + 0.2 * exposure_score
return float(np.clip(quality, 0.0, 1.0))
# =============================================================================
# Master Step 2 Function
# =============================================================================
def step2_illumination(
step1_result: NormalizedEyePatch,
iris_radius_frame: float,
) -> IlluminationResult:
patch_in = step1_result.norm_crop
deglinted, glint_mask = remove_glints_adaptive(patch_in)
iris_r_patch = estimate_iris_radius_in_patch(
iris_radius_frame, step1_result.warp_M
)
iris_mask = build_iris_mask(
iris_center_patch=step1_result.iris_center,
iris_radius_patch=iris_r_patch,
patch_shape=(PATCH_H, PATCH_W),
feather_width=0.35,
)
sclera_mask = 1.0 - iris_mask
iris_clahe, sclera_clahe, blended = region_aware_clahe(deglinted, iris_mask)
global_clahe = CLAHE_GLOBAL.apply(deglinted)
itv = iris_texture_visibility(blended, iris_mask)
photo_q = photometric_quality(blended, iris_mask, glint_mask)
is_usable = itv > ITV_THRESHOLD and step1_result.is_open
diff_vs_global = cv2.applyColorMap(
cv2.convertScaleAbs(cv2.absdiff(blended, global_clahe), alpha=4.0),
cv2.COLORMAP_PLASMA
)
diff_vs_raw = cv2.applyColorMap(
cv2.convertScaleAbs(cv2.absdiff(blended, patch_in), alpha=4.0),
cv2.COLORMAP_INFERNO
)
return IlluminationResult(
step1=step1_result,
iris_mask=iris_mask,
sclera_mask=sclera_mask,
iris_clahe=iris_clahe,
sclera_clahe=sclera_clahe,
blended=blended,
global_clahe=global_clahe,
diff_vs_global=diff_vs_global,
diff_vs_raw=diff_vs_raw,
itv=itv,
photo_quality=photo_q,
is_usable=is_usable,
)
# =============================================================================
# STEP 3 - Limbus Circle Fitting (Seeded Radial Edge Profiling)
# =============================================================================
from dataclasses import field
from typing import List
NUM_RAYS = 36
RAY_SEARCH_BAND = 0.45
MIN_VALID_RAYS = 10
RANSAC_ITERS = 120
RANSAC_THRESH = 1.8
@dataclass
class LimbusResult:
step2: "IlluminationResult"
ellipse: Optional[tuple]
edge_points: np.ndarray
ray_valid: np.ndarray
inlier_mask: np.ndarray
limbus_center: Optional[Tuple[float,float]]
limbus_axes: Optional[Tuple[float,float]]
limbus_angle: float
eccentricity: float
pupil_center: Optional[Tuple[float,float]]
pupil_radius: float
angular_coverage: float
fit_residual: float
limbus_quality: float
is_reliable: bool
def sample_radial_rays(
patch: np.ndarray,
center: Tuple[float, float],
seed_radius: float,
n_rays: int = NUM_RAYS,
band: float = RAY_SEARCH_BAND,
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
H, W = patch.shape[:2]
cx, cy = center
r_min = max(1.0, seed_radius * (1.0 - band))
r_max = min(seed_radius * (1.0 + band),
min(cx, cy, W - cx, H - cy) - 1)
r_max = max(r_max, r_min + 2.0)
n_samples = max(12, int((r_max - r_min) * 3))
r_coords = np.linspace(r_min, r_max, n_samples, dtype=np.float32)
angles = np.linspace(0, 2 * np.pi, n_rays, endpoint=False,
dtype=np.float32)
profiles = np.zeros((n_rays, n_samples), dtype=np.float32)
for i, theta in enumerate(angles):
cos_t = np.cos(theta)
sin_t = np.sin(theta)
xs = np.clip(cx + r_coords * cos_t, 0, W - 1).astype(np.float32)
ys = np.clip(cy + r_coords * sin_t, 0, H - 1).astype(np.float32)
profiles[i] = cv2.remap(
patch.astype(np.float32),
xs.reshape(1, -1), ys.reshape(1, -1),
cv2.INTER_LINEAR
).ravel()
return profiles, r_coords, angles
def find_limbus_edge_on_ray(
profile: np.ndarray,
r_coords: np.ndarray,
expected_r: float,
) -> Tuple[Optional[float], float]:
if len(profile) < 4:
return None, 0.0
profile_s = cv2.GaussianBlur(
profile.reshape(1, -1).astype(np.float32),
(1, 5), sigmaX=1.0
).ravel()
grad = np.gradient(profile_s.astype(np.float64))
best_idx = int(np.argmax(grad))
strength = float(grad[best_idx])
if strength < 2.0:
return None, 0.0
edge_r = float(r_coords[best_idx])
return edge_r, strength
def extract_edge_points(
profiles: np.ndarray,
r_coords: np.ndarray,
angles: np.ndarray,
center: Tuple[float, float],
seed_radius: float,
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
"""
PATCHED: Fixed bounds check to use PATCH_W and PATCH_H instead of profiles.shape[0]
"""
cx, cy = center
edge_pts = []
strengths = []
ray_valid = np.zeros(len(angles), dtype=bool)
for i, (theta, profile) in enumerate(zip(angles, profiles)):
edge_r, strength = find_limbus_edge_on_ray(profile, r_coords, seed_radius)
if edge_r is not None:
x = cx + edge_r * np.cos(theta)
y = cy + edge_r * np.sin(theta)
# PATCHED: Use PATCH_W and PATCH_H instead of profiles.shape[0]
if 0 <= x < PATCH_W and 0 <= y < PATCH_H:
edge_pts.append([x, y])
strengths.append(strength)
ray_valid[i] = True
if len(edge_pts) == 0:
return np.zeros((0, 2), dtype=np.float32), ray_valid, np.zeros(0)
return (np.array(edge_pts, dtype=np.float32),
ray_valid,
np.array(strengths, dtype=np.float32))
def fit_ellipse_ransac(
points: np.ndarray,
n_iters: int = RANSAC_ITERS,
thresh: float = RANSAC_THRESH,
) -> Tuple[Optional[tuple], np.ndarray]:
N = len(points)
if N < 6:
return None, np.zeros(N, dtype=bool)
if N < 10:
try:
ell = cv2.fitEllipse(points)
return ell, np.ones(N, dtype=bool)
except cv2.error:
return None, np.zeros(N, dtype=bool)
best_ellipse = None
best_inliers = np.zeros(N, dtype=bool)
best_n_inliers = 0
rng = np.random.default_rng(42)
for _ in range(n_iters):
idx = rng.choice(N, 5, replace=False)
sample = points[idx]
try:
ell = cv2.fitEllipse(sample)
except cv2.error:
continue
(cx, cy), (ma, mi), angle = ell
if ma < 1 or mi < 1 or ma > 200 or mi > 200:
continue
if ma / (mi + 1e-6) > 4.0:
continue
inliers = _ellipse_inliers(points, ell, thresh)
n_in = int(inliers.sum())
if n_in > best_n_inliers:
best_n_inliers = n_in
best_inliers = inliers
best_ellipse = ell
if best_ellipse is not None and best_n_inliers >= 5:
try:
best_ellipse = cv2.fitEllipse(points[best_inliers])
except cv2.error:
pass
return best_ellipse, best_inliers
def _ellipse_inliers(
points: np.ndarray,
ellipse: tuple,
thresh: float,
) -> np.ndarray:
(cx, cy), (ma, mi), angle_deg = ellipse
a = np.radians(angle_deg)
cos_a, sin_a = np.cos(a), np.sin(a)
dx = points[:, 0] - cx
dy = points[:, 1] - cy
u = dx * cos_a + dy * sin_a
v = -dx * sin_a + dy * cos_a
ra = ma / 2.0 + 1e-6
rb = mi / 2.0 + 1e-6
f = (u / ra)**2 + (v / rb)**2
dist = np.abs(f - 1.0) * (ra * rb) / (ra + rb)
return dist < thresh
def estimate_pupil_seed(
patch: np.ndarray,
limbus_ellipse: Optional[tuple],
iris_center: Tuple[float, float],
iris_radius: float,
) -> Tuple[Optional[Tuple[float,float]], float]:
H, W = patch.shape[:2]
mask = np.zeros((H, W), dtype=np.uint8)
if limbus_ellipse is not None:
cv2.ellipse(mask, limbus_ellipse, 255, -1)
else:
cv2.circle(mask, (int(iris_center[0]), int(iris_center[1])),
int(iris_radius * 0.9), 255, -1)
if cv2.countNonZero(mask) < 5:
return None, 0.0
masked = cv2.bitwise_and(patch, patch, mask=mask)
iris_pixels = patch[mask > 0]
if len(iris_pixels) < 5:
return None, 0.0
dark_thresh = float(np.percentile(iris_pixels, 25))
_, dark_mask = cv2.threshold(masked, int(dark_thresh), 255,
cv2.THRESH_BINARY_INV)
dark_mask = cv2.bitwise_and(dark_mask, mask)
k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
dark_mask = cv2.morphologyEx(dark_mask, cv2.MORPH_OPEN, k, iterations=1)
contours, _ = cv2.findContours(dark_mask, cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
if not contours:
return None, 0.0
best = max(contours, key=cv2.contourArea)
area = cv2.contourArea(best)
if area < 4:
return None, 0.0
M = cv2.moments(best)
if M["m00"] < 1e-6:
return None, 0.0
pcx = M["m10"] / M["m00"]
pcy = M["m01"] / M["m00"]
pradius = float(np.sqrt(area / np.pi))
return (float(pcx), float(pcy)), pradius
def compute_limbus_quality(
ellipse: Optional[tuple],
inlier_mask: np.ndarray,
edge_points: np.ndarray,
ray_valid: np.ndarray,
iris_seed_radius: float,
) -> Tuple[float, float, float]:
angular_coverage = float(ray_valid.sum()) / max(len(ray_valid), 1)
if ellipse is None or inlier_mask.sum() < 5:
return angular_coverage, 999.0, 0.0
inlier_pts = edge_points[inlier_mask]
dists = _point_to_ellipse_dist(inlier_pts, ellipse)
fit_resid = float(np.mean(dists)) if len(dists) > 0 else 999.0
(cx, cy), (ma, mi), _ = ellipse
ratio = min(ma, mi) / (max(ma, mi) + 1e-6)
size_ok = 0.5 < (ma / 2) / (iris_seed_radius + 1e-6) < 2.0
coverage_score = angular_coverage
residual_score = float(np.clip(1.0 - fit_resid / 5.0, 0.0, 1.0))
shape_score = ratio * (1.0 if size_ok else 0.3)
quality = 0.4 * coverage_score + 0.4 * residual_score + 0.2 * shape_score
return angular_coverage, fit_resid, float(np.clip(quality, 0.0, 1.0))
def _point_to_ellipse_dist(
points: np.ndarray,
ellipse: tuple,
) -> np.ndarray:
if len(points) == 0:
return np.zeros(0)
(cx, cy), (ma, mi), angle_deg = ellipse
a = np.radians(angle_deg)
ca, sa = np.cos(a), np.sin(a)
dx, dy = points[:,0] - cx, points[:,1] - cy
u = dx*ca + dy*sa
v = -dx*sa + dy*ca
ra, rb = ma/2.0 + 1e-6, mi/2.0 + 1e-6
f = (u/ra)**2 + (v/rb)**2
return np.abs(f - 1.0) * (ra*rb)/(ra+rb)
def step3_limbus(
step2_result: "IlluminationResult",
iris_radius_frame: float,
) -> LimbusResult:
patch = step2_result.blended
iris_c = step2_result.step1.iris_center
warp_M = step2_result.step1.warp_M
scale = float(np.linalg.norm(warp_M[0, :2]))
iris_r_px = float(np.clip(iris_radius_frame * scale, 3.0, min(PATCH_W, PATCH_H) / 2.0 - 1))
profiles, r_coords, angles = sample_radial_rays(
patch, iris_c, iris_r_px,
n_rays=NUM_RAYS, band=RAY_SEARCH_BAND
)
edge_pts, ray_valid, strengths = extract_edge_points(
profiles, r_coords, angles, iris_c, iris_r_px
)
if len(edge_pts) >= MIN_VALID_RAYS:
ellipse, inlier_mask = fit_ellipse_ransac(edge_pts)
else:
ellipse = None
inlier_mask = np.zeros(len(edge_pts), dtype=bool)
if ellipse is None:
cx, cy = iris_c
ellipse = (
(float(cx), float(cy)),
(float(iris_r_px * 2), float(iris_r_px * 2)),
0.0
)
inlier_mask = np.ones(len(edge_pts), dtype=bool)
(ecx, ecy), (ma, mi), eangle = ellipse
limbus_center = (float(ecx), float(ecy))
limbus_axes = (float(ma / 2.0), float(mi / 2.0))
ecc = float(np.sqrt(max(0, 1.0 - (min(ma,mi)/max(ma,mi,1e-6))**2)))
pupil_center, pupil_radius = estimate_pupil_seed(
patch, ellipse, iris_c, iris_r_px
)
angular_cov, fit_resid, lq = compute_limbus_quality(
ellipse, inlier_mask, edge_pts, ray_valid, iris_r_px
)
is_reliable = lq > 0.35 and angular_cov > 0.4
return LimbusResult(
step2=step2_result,
ellipse=ellipse,
edge_points=edge_pts,
ray_valid=ray_valid,
inlier_mask=inlier_mask,
limbus_center=limbus_center,
limbus_axes=limbus_axes,
limbus_angle=float(eangle),
eccentricity=ecc,
pupil_center=pupil_center,
pupil_radius=pupil_radius,
angular_coverage=angular_cov,
fit_residual=fit_resid,
limbus_quality=lq,
is_reliable=is_reliable,
)
if __name__ == "__main__":
import argparse
print("preprocessing_pipeline.py loaded OK - this module is meant to be imported,")
print("not run directly, by calibrate.py / run_session.py / main_webcam_pipeline.py.")
|