Corner detection (Harris) in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - Corner detection (Harris)

Which metric matters for Corner Detection (Harris) and WHY

For corner detection, the key metrics are repeatability and localization accuracy. Repeatability means the detector finds the same corners under different views or lighting. Localization accuracy means the detected corners are close to the true corner points. These metrics matter because a good corner detector should reliably find important points that help in tasks like image matching or tracking.

Confusion matrix or equivalent visualization

True Corners Detected: TP = 85
False Corners Detected: FP = 15
Missed True Corners: FN = 10
Correctly Ignored Non-Corners: TN = 890

Confusion Matrix:
          | Detected Corner | Not Detected |
----------|-----------------|--------------|
Corner    |       85 (TP)   |    10 (FN)   |
No Corner |       15 (FP)   |   890 (TN)   |

Total samples = 85 + 15 + 10 + 890 = 1000
    

Precision vs Recall Tradeoff with Examples

Precision tells us how many detected corners are actually true corners. High precision means few false corners, which is good to avoid noise.

Recall tells us how many true corners were detected. High recall means we miss very few real corners.

For example, if you want to track objects in video, missing corners (low recall) can cause tracking failure. But if you detect too many false corners (low precision), the system wastes time processing useless points.

So, a balance is needed. Harris detector parameters can be tuned to increase recall (detect more corners) or precision (detect fewer but more accurate corners).

What "Good" vs "Bad" Metric Values Look Like

Good: Precision and recall both above 0.8 means most detected corners are true and most true corners are found. Localization error is low (corners detected within a few pixels of true corners).

Bad: Precision below 0.5 means many false corners detected, causing noise. Recall below 0.5 means many true corners missed, losing important features. Large localization error means corners are not accurately placed.

Common Pitfalls in Metrics for Corner Detection

• Ignoring localization error: Counting detected corners without checking how close they are to true corners can mislead about quality.
• Overfitting to one image: A detector tuned only for one image may not work well on others, hurting repeatability.
• Data leakage: Using test images to tune parameters inflates performance metrics.
• Accuracy paradox: High accuracy can be misleading if most pixels are non-corners (TN dominate), so precision and recall are better.

Self-Check Question

Your Harris corner detector has 98% accuracy but only 12% recall on true corners. Is it good for use?

Answer: No, because the detector misses 88% of true corners (low recall). High accuracy is misleading here since most pixels are non-corners. The detector is not reliable for finding important corners.

Key Result

For Harris corner detection, high recall and precision with low localization error ensure reliable and accurate corner points.

Practice

1. What is the main goal of the Harris corner detection algorithm in computer vision?

easy

A. To detect straight lines in an image

B. To find points in an image where edges meet, called corners

C. To blur the image for noise reduction

D. To segment the image into different color regions

Solution

1. Step 1: Understand the purpose of Harris corner detection

   Harris corner detection is designed to find corners, which are points where two edges meet in an image.

2. Step 2: Compare with other options

   Blurring, line detection, and segmentation are different tasks not performed by Harris corner detection.

3. Final Answer:

   To find points in an image where edges meet, called corners -> Option B

4. Quick Check:

   Harris detects corners = C [OK]

Hint: Corners are where edges meet, Harris finds these points [OK]

Common Mistakes:

• Confusing corner detection with edge detection
• Thinking Harris blurs or segments images
• Mixing up line detection with corner detection

2. Which of the following is the correct formula for the Harris corner response R?

easy

A. R = det(M) + k * (trace(M))^2

B. R = trace(M) - k * det(M)

C. R = det(M) - k * (trace(M))^2

D. R = det(M) / trace(M)

Solution

1. Step 1: Recall the Harris corner response formula

   The Harris response is calculated as R = det(M) - k * (trace(M))^2, where M is the second moment matrix and k is a sensitivity factor.

2. Step 2: Verify other options

   Other formulas either add instead of subtract or mix det and trace incorrectly.

3. Final Answer:

   R = det(M) - k * (trace(M))^2 -> Option C

4. Quick Check:

   Harris R formula uses det minus k times trace squared [OK]

Hint: Remember: R = det minus k times trace squared [OK]

Common Mistakes:

• Adding instead of subtracting in the formula
• Confusing determinant with trace
• Using division instead of subtraction

3. Given the following Python code snippet using OpenCV, what will be the output type of corners?

import cv2
import numpy as np
img = cv2.imread('image.jpg', 0)
corners = cv2.cornerHarris(img, 2, 3, 0.04)
print(type(corners))

medium

A. <class 'float'>

B. <class 'list'>

C. <class 'int'>

D. <class 'numpy.ndarray'>

Solution

1. Step 1: Understand OpenCV cornerHarris output

   The function cv2.cornerHarris returns a numpy array representing the corner response for each pixel.

2. Step 2: Check the printed type

   Printing type(corners) will show <class 'numpy.ndarray'> because corners is a numpy array.

3. Final Answer:

   <class 'numpy.ndarray'> -> Option D

4. Quick Check:

   cornerHarris returns numpy array [OK]

Hint: cornerHarris returns a numpy array of responses [OK]

Common Mistakes:

• Assuming output is a list instead of numpy array
• Thinking output is a single number
• Confusing output type with image type

4. In the code below, why does the Harris corner detection not highlight any corners?

import cv2
import numpy as np
img = cv2.imread('image.jpg', 0)
corners = cv2.cornerHarris(img, 2, 3, 0.04)
corners = cv2.dilate(corners, None)
img[corners > 0.01 * corners.max()] = 255
cv2.imshow('Corners', img)
cv2.waitKey(0)
cv2.destroyAllWindows()

medium

A. The dilation step is missing a kernel argument

B. The threshold 0.01 * corners.max() is too high, no pixels pass

C. The image is grayscale, so corners cannot be detected

D. The assignment img[corners > threshold] = 255 modifies the original image incorrectly

Solution

1. Step 1: Analyze the dilation step

   The line corners = cv2.dilate(corners, None) fails with a TypeError because cv2.dilate requires a kernel (e.g., np.ones((3,3), np.uint8)). None is invalid, causing the code to crash before imshow.

2. Step 2: Rule out other options

   Grayscale works fine (A wrong), 0.01 threshold is standard (B wrong), direct modification to 255 is common for grayscale marking (D ok).

3. Final Answer:

   The dilation step is missing a kernel argument -> Option A

4. Quick Check:

   cv2.dilate requires kernel [OK]

Hint: cv2.dilate needs kernel like np.ones((3,3)); None causes TypeError [OK]

Common Mistakes:

• Thinking grayscale images can't have corners
• Assuming threshold is too high
• Believing img modification is incorrect (common for grayscale)

5. You want to detect strong corners in a noisy image using Harris corner detection. Which combination of steps will best improve corner detection accuracy?

hard

A. Apply Gaussian blur before detection, use a moderate window size, and set a proper threshold to filter weak corners

B. Use raw noisy image, large window size, and low threshold for more corners

C. Apply Gaussian blur before detection, use a smaller window size, and increase threshold

D. Skip blurring, use smallest window size, and no threshold to detect all corners

Solution

1. Step 1: Understand noise impact and preprocessing

   Noise can cause false corners, so applying Gaussian blur smooths the image and reduces noise effects.

2. Step 2: Choose window size and threshold carefully

   A moderate window size balances detail and noise, and a proper threshold filters out weak corners, improving accuracy.

3. Final Answer:

   Apply Gaussian blur before detection, use a moderate window size, and set a proper threshold to filter weak corners -> Option A

4. Quick Check:

   Blur + moderate window + threshold = better corners [OK]

Hint: Blur first, then moderate window and threshold for best corners [OK]

Common Mistakes:

• Ignoring noise and skipping blur
• Using too small or too large window size
• Setting threshold too low or too high