DNN-based face detection in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - DNN-based face detection

Which metric matters for DNN-based face detection and WHY

For face detection using deep neural networks, the key metrics are Precision and Recall. Precision tells us how many detected faces are actually faces, so it measures false alarms. Recall tells us how many real faces the model finds, so it measures missed faces. Both matter because we want to find as many faces as possible (high recall) without wrongly marking non-faces as faces (high precision). The F1 score balances these two. Also, the Average Precision (AP) over different confidence thresholds is often used to summarize performance.

Confusion matrix example for face detection

      | Predicted Face | Predicted No Face |
      |---------------|-------------------|
      | True Face (TP)  90                  10 (FN)          |
      | False Face (FP) 15                  885 (TN)         |

      Total samples = 1000

      Precision = TP / (TP + FP) = 90 / (90 + 15) = 0.857
      Recall = TP / (TP + FN) = 90 / (90 + 10) = 0.9
      F1 Score = 2 * (0.857 * 0.9) / (0.857 + 0.9) ≈ 0.878
    

Precision vs Recall tradeoff with examples

If the model is tuned to be very strict, it will only detect faces when very sure. This means high precision (few false alarms) but low recall (misses many faces). This is good if false alarms are costly, like in security checks.

If the model is tuned to detect as many faces as possible, it will catch almost all faces (high recall) but also mark some non-faces as faces (low precision). This is useful in photo apps where missing a face is worse than a few false detections.

Choosing the right balance depends on the application needs.

What good vs bad metric values look like for face detection

• Good: Precision and Recall both above 0.85, F1 score near 0.9 or higher. This means most faces are found and few false alarms.
• Bad: Precision or Recall below 0.5 means many false alarms or many missed faces. For example, Precision 0.4 means more than half detected faces are wrong.
• Very high accuracy alone can be misleading if the dataset has many non-face images (class imbalance).

Common pitfalls in evaluating face detection metrics

• Accuracy paradox: If most images have no faces, a model that always predicts no face can have high accuracy but is useless.
• Data leakage: Testing on images very similar to training images inflates metrics falsely.
• Overfitting: Very high training metrics but low test metrics means the model memorizes training faces but fails on new ones.
• Ignoring IoU thresholds: Face detection usually requires bounding boxes to overlap enough with ground truth. Metrics must consider this overlap.

Self-check question

Your face detection model has 98% accuracy but only 12% recall on faces. Is it good for production? Why or why not?

Answer: No, it is not good. The high accuracy is misleading because most images may not have faces. The very low recall means the model misses 88% of faces, which defeats the purpose of face detection.

Key Result

Precision and recall are key for face detection; balance them to avoid missing faces or false alarms.

Practice

1. What is the main purpose of using a DNN-based face detection model?

easy

A. To increase the size of an image

B. To find faces in images automatically

C. To compress images for storage

D. To change the color of an image

Solution

1. Step 1: Understand the goal of face detection

   Face detection means finding where faces are in pictures or videos.

2. Step 2: Recognize what DNN-based models do

   DNN models learn patterns to automatically locate faces without manual rules.

3. Final Answer:

   To find faces in images automatically -> Option B

4. Quick Check:

   Face detection = find faces [OK]

Hint: Face detection means locating faces in images [OK]

Common Mistakes:

• Confusing detection with image editing
• Thinking it changes image colors
• Mixing detection with compression

2. Which OpenCV function is used to load a pre-trained DNN face detection model?

easy

A. cv2.dnn.readNet()

B. cv2.imread()

C. cv2.resize()

D. cv2.cvtColor()

Solution

1. Step 1: Identify the function for loading DNN models

   OpenCV's DNN module uses cv2.dnn.readNet() to load models from files.

2. Step 2: Differentiate from other functions

   cv2.imread() loads images, cv2.resize() changes size, cv2.cvtColor() changes color space.

3. Final Answer:

   cv2.dnn.readNet() -> Option A

4. Quick Check:

   Load DNN model = cv2.dnn.readNet() [OK]

Hint: Use cv2.dnn.readNet() to load DNN models [OK]

Common Mistakes:

• Using cv2.imread() to load models
• Confusing image functions with model loading
• Trying to resize models with cv2.resize()

3. Given this code snippet, what will be the output type of detections?

net = cv2.dnn.readNet('model.caffemodel', 'deploy.prototxt')
blob = cv2.dnn.blobFromImage(image, 1.0, (300, 300), (104, 177, 123))
net.setInput(blob)
detections = net.forward()

medium

A. A list of bounding box coordinates

B. A single integer count of faces

C. A string describing detected faces

D. A NumPy array with detection data

Solution

1. Step 1: Understand net.forward() output

   net.forward() returns a NumPy array containing detection results with scores and box info.

2. Step 2: Differentiate output types

   It is not a list, integer, or string but a multi-dimensional array with detection details.

3. Final Answer:

   A NumPy array with detection data -> Option D

4. Quick Check:

   net.forward() output = NumPy array [OK]

Hint: net.forward() returns a NumPy array of detections [OK]

Common Mistakes:

• Thinking output is a list or count
• Expecting a string description
• Confusing output with input image

4. You run this code but get no faces detected. What is the likely error?

blob = cv2.dnn.blobFromImage(image, 1.0, (224, 224), (104, 177, 123))
net.setInput(blob)
detections = net.forward()

What should you fix?

medium

A. Change input size to (300, 300) to match model requirements

B. Remove mean subtraction values from blobFromImage

C. Use cv2.imread() instead of blobFromImage

D. Call net.forward() before setInput()

Solution

1. Step 1: Check model input size requirements

   Many DNN face detectors expect input size (300, 300). Using (224, 224) causes mismatch and no detection.

2. Step 2: Correct the input size in blobFromImage

   Changing to (300, 300) matches model and enables detection.

3. Final Answer:

   Change input size to (300, 300) to match model requirements -> Option A

4. Quick Check:

   Input size must match model = (300, 300) [OK]

Hint: Match blob size to model input size [OK]

Common Mistakes:

• Using wrong input size for the model
• Calling net.forward() before setInput()
• Removing mean subtraction incorrectly

5. You want to improve face detection accuracy on images with multiple faces at different sizes. Which approach helps most?

hard

A. Reduce the input image size to speed up detection

B. Increase the learning rate during training

C. Use image pyramid to detect faces at multiple scales

D. Remove mean subtraction from preprocessing

Solution

1. Step 1: Understand multi-scale detection challenge

   Faces appear at different sizes; detecting all requires checking multiple scales.

2. Step 2: Use image pyramid technique

   Image pyramid resizes image multiple times to detect faces at various sizes, improving accuracy.

3. Final Answer:

   Use image pyramid to detect faces at multiple scales -> Option C

4. Quick Check:

   Multi-scale detection = image pyramid [OK]

Hint: Use image pyramid for faces at different sizes [OK]

Common Mistakes:

• Thinking higher learning rate fixes scale issues
• Reducing image size loses small faces
• Skipping mean subtraction harms accuracy