DNN-based face detection in Computer Vision - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding the role of anchor boxes in DNN face detection

In deep neural network (DNN) based face detection, what is the primary purpose of anchor boxes?

A To generate synthetic face images for training the model

B To reduce the number of layers in the neural network for faster computation

C To provide predefined bounding boxes at different scales and aspect ratios to help detect faces of various sizes

D To normalize the input images before feeding them into the network

💡 Hint

Think about how the model predicts where faces might be in an image.

Anchor boxes are predefined boxes of various sizes and shapes that help the model predict bounding boxes for faces at different scales and aspect ratios.

❓ Predict Output

intermediate

Output shape of the final layer in a face detection DNN

Consider a DNN face detector that outputs a tensor of shape (batch_size, 8732, 7). What does the number 7 represent in this output?

A It represents 7 different classes of objects detected including faces

B It represents 4 coordinates for bounding box, 2 for class scores, and 1 for confidence score

C It represents 7 different face landmarks detected per box

D It represents 4 coordinates for bounding box, 2 for class probabilities (face or background), and 1 for an objectness score

💡 Hint

Recall that face detection usually involves bounding box coordinates and classification between face and background.

The 7 values typically include 4 for bounding box coordinates, 2 for class probabilities (face vs background), and 1 objectness score indicating confidence.

❓ Hyperparameter

advanced

Choosing the right confidence threshold in face detection

In a DNN-based face detector, what is the effect of setting a very high confidence threshold for detected faces?

A It reduces false positives but may increase false negatives by missing some faces

B It increases both false positives and false negatives

C It increases true positives but also increases false positives

D It has no effect on detection results

💡 Hint

Think about what happens if the model only accepts very confident detections.

A high confidence threshold means the model only accepts detections it is very sure about, reducing false alarms but possibly missing some real faces.

❓ Metrics

advanced

Evaluating face detection with Intersection over Union (IoU)

Which statement best describes how Intersection over Union (IoU) is used to evaluate face detection models?

A IoU is used to measure the confidence score of detected faces

B IoU measures the overlap between predicted and ground truth bounding boxes to determine if a detection is correct

C IoU calculates the pixel-wise difference between input images and detected faces

D IoU counts the total number of faces detected in an image

💡 Hint

Think about how to check if a predicted box matches the actual face location.

IoU calculates the ratio of the overlapping area to the combined area of predicted and true bounding boxes, helping decide if a detection is accurate.

🔧 Debug

expert

Debugging low face detection accuracy due to input preprocessing

A DNN face detection model trained on RGB images shows very low accuracy when tested on new images. The new images are loaded as grayscale and then converted to 3 channels by repeating the grayscale channel. What is the most likely cause of the low accuracy?

A The model expects color features from RGB images, but repeated grayscale channels lack color information, causing poor detection

B The model requires images to be in grayscale, so RGB input causes low accuracy

C Grayscale images have different pixel value ranges, causing normalization errors in the model

D Repeating grayscale channels increases image size, causing memory errors during inference

💡 Hint

Consider how color information affects feature extraction in CNNs.

The model was trained on RGB images with color features. Repeating grayscale channels creates fake RGB input without true color, leading to poor feature extraction and low accuracy.