#!/usr/bin/env python3 """ Compute per-file IoU and mean-IoU between an *inverted* reference mask and several *inverted* transformed risk-maps, visualised as a colour overlay. USAGE ----- python compute_miou.py MASK_PATH COORDS_PATH [--delay 400] MASK_PATH : reference mask image COORDS_PATH : coordinates.txt (filename angle scale offset_x offset_y) --delay : milliseconds each overlay is shown (0 = wait for key) """ import argparse import sys from pathlib import Path import cv2 import numpy as np from utils import affine, parse_coords_path # ────────── CLI ──────────────────────────────────────────────────────────────── ap = argparse.ArgumentParser(description="mIoU (inverted masks) with overlay") ap.add_argument("mask", help="reference mask image") ap.add_argument("coords", help="coordinates.txt file") ap.add_argument( "--delay", type=int, default=400, help="delay per overlay in ms (0 = wait)" ) args = ap.parse_args() delay = max(args.delay, 0) mask_path = Path(args.mask).expanduser().resolve() coords_path = Path(args.coords).expanduser().resolve() risk_root = coords_path.parent # ────────── helpers ──────────────────────────────────────────────────────────── def binarise_and_invert(img, thr=0.99): """gray/RGB → binary (0/1) then invert (1-p).""" if img.ndim == 3: img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) bin_img = (img.astype(np.float32) / 255.0 > thr).astype(np.uint8) return 1 - bin_img # invert here! def iou(a, b): """Intersection over Union (IoU) = TP / (TP + FP + FN)""" inter = np.logical_and(a, b).sum() union = np.logical_or(a, b).sum() return inter / union if union else 1.0 def recall(a, b): """Recall = TP / (TP + FN)""" tp = np.logical_and(a, b).sum() fn = np.logical_and(a, np.logical_not(b)).sum() return tp / (tp + fn) if (tp + fn) else 1.0 def precision(a, b): """Precision = TP / (TP + FP)""" tp = np.logical_and(a, b).sum() fp = np.logical_and(np.logical_not(a), b).sum() return tp / (tp + fp) if (tp + fp) else 1.0 def f1_score(a, b): """F1 Score = 2 * (Precision * Recall) / (Precision + Recall)""" p = precision(a, b) r = recall(a, b) return 2 * p * r / (p + r) if (p + r) else 0.0 # ────────── load & invert reference mask ─────────────────────────────────────── mask_img = cv2.imread(str(mask_path), cv2.IMREAD_UNCHANGED) if mask_img is None: sys.exit(f"Error: cannot open {mask_path}") mask_full = binarise_and_invert(mask_img) H_full, W_full = mask_full.shape # ────────── parse the coordinate path file ───────────────────────────────────── files = parse_coords_path(coords_path) # ────────── main loop ────────────────────────────────────────────────────────── ious = {} precisions = {} recalls = {} f1_scores = {} cv2.namedWindow("overlay", cv2.WINDOW_NORMAL) for k, v in files.items(): fname = k ang = v["ang"] scl = v["scl"] tx = v["tx"] ty = v["ty"] print(fname, ang, scl, tx, ty) ang, scl, tx, ty = map(float, (ang, scl, tx, ty)) risk_path = Path(fname) if Path(fname).is_absolute() else risk_root / fname risk_img = cv2.imread(str(risk_path), cv2.IMREAD_UNCHANGED) h, w = risk_img.shape[:2] print(f"Processing {risk_path} ({h}x{w})") # take the middle N/2 x N/2 pixels if the image risk_img = risk_img[h // 4 : 3 * h // 4, w // 4 : 3 * w // 4] h, w = risk_img.shape[:2] # # adjust the affine transform to match the new size tx = tx + w // 2 ty = ty + h // 2 if risk_img is None: print(f"[warning] cannot read {risk_path}; skipping.") continue risk_bin = binarise_and_invert(risk_img) M = affine(ang, scl, tx, ty, w, h) warped = cv2.warpAffine( risk_bin, M, (W_full, H_full), flags=cv2.INTER_NEAREST, borderValue=0 ) # ─── crop to bounding box of warped risk map ─────────────────────────── nz = cv2.findNonZero(warped) if nz is None: x, y, w_box, h_box = 0, 0, W_full, H_full else: x, y, w_box, h_box = cv2.boundingRect(nz) pad = 10 x = max(x - pad, 0) y = max(y - pad, 0) w_box = min(w_box + 2 * pad, W_full - x) h_box = min(h_box + 2 * pad, H_full - y) mask_crop = mask_full[y : y + h_box, x : x + w_box] warped_crop = warped[y : y + h_box, x : x + w_box] this_iou = iou(mask_crop, warped_crop) this_recall = recall(mask_crop, warped_crop) this_precision = precision(mask_crop, warped_crop) this_f1_score = f1_score(mask_crop, warped_crop) ious[risk_path] = this_iou precisions[risk_path] = this_precision recalls[risk_path] = this_recall f1_scores[risk_path] = this_f1_score # ─── colour overlay (green = inverted mask, red = inverted risk) ────── overlay = np.zeros((h_box, w_box, 3), np.uint8) overlay[..., 1] = mask_crop * 255 overlay[..., 2] = warped_crop * 255 cv2.imshow("overlay", overlay) cv2.setWindowTitle("overlay", f"{risk_path.name} | IoU = {this_iou:.4f}") k = cv2.waitKey(delay) & 0xFF if k in (27, ord("q"), ord("Q")): break cv2.destroyAllWindows() # ────────── results ──────────────────────────────────────────────────────────── if ious: print("\nPer-file IoUs:") for k, val in ious.items(): path = Path(k) last_path = path.parent.name + "/" + path.name print(f" # {last_path}: {val:.4f}") print(f"\nMean IoU over {len(ious)} files: {np.mean(list(ious.values())):.4f}") else: print("No risk-map files processed for ious.") if precisions: print("\nPer-file Precisions:") for k, val in precisions.items(): path = Path(k) last_path = path.parent.name + "/" + path.name print(f" # {last_path}: {val:.4f}") print(f"\nMean Precision over {len(precisions)} files: {np.mean(list(precisions.values())):.4f}") else: print("No risk-map files processed for precision.") if recalls: print("\nPer-file Recalls:") for k, val in recalls.items(): path = Path(k) last_path = path.parent.name + "/" + path.name print(f" # {last_path}: {val:.4f}") print(f"\nMean Recall over {len(recalls)} files: {np.mean(list(recalls.values())):.4f}") else: print("No risk-map files processed for recall.") if f1_scores: print("\nPer-file F1 Scores:") for k, val in f1_scores.items(): path = Path(k) last_path = path.parent.name + "/" + path.name print(f" # {last_path}: {val:.4f}") print(f"\nMean F1 Score over {len(f1_scores)} files: {np.mean(list(f1_scores.values())):.4f}") else: print("No risk-map files processed for F1 score.")