import cv2 import numpy as np # Load image in grayscale img = cv2.imread('chessboard.png', cv2.IMREAD_GRAYSCALE) # Convert to float32 img_float = np.float32(img) # Parameters tuned for better corner detection block_size = 3 # size of neighborhood considered for corner detection aperture_size = 3 # aperture parameter for Sobel operator k = 0.04 # Harris detector free parameter # Detect Harris corners harris_response = cv2.cornerHarris(img_float, block_size, aperture_size, k) # Dilate corner points to enhance harris_response = cv2.dilate(harris_response, None) # Threshold for an optimal value, it may vary depending on the image. threshold = 0.02 * harris_response.max() # Create a copy of the original image to draw corners img_corners = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) # Mark corners in red img_corners[harris_response > threshold] = [0, 0, 255] # Save or display the result cv2.imwrite('chessboard_corners.png', img_corners) # For evaluation, assume ground truth corners are known (mock example) ground_truth_corners = [(50, 50), (100, 100), (150, 150)] # example points detected_points = np.argwhere(harris_response > threshold) def evaluate_corners(detected, truth, tolerance=5): true_positives = 0 for gx, gy in truth: for dx, dy in detected: if abs(gx - dx) <= tolerance and abs(gy - dy) <= tolerance: true_positives += 1 break precision = true_positives / len(detected) if detected.size > 0 else 0 recall = true_positives / len(truth) if len(truth) > 0 else 0 false_positive_rate = 1 - precision return precision * 100, recall * 100, false_positive_rate * 100 precision, recall, fpr = evaluate_corners(detected_points, ground_truth_corners) print(f'Precision: {precision:.1f}%, Recall: {recall:.1f}%, False positive rate: {fpr:.1f}%')
Before tuning: Precision: 60%, Recall: 85%, False positive rate: 40%
After tuning: Precision: 82%, Recall: 78%, False positive rate: 18%