Machine Learning Computer Vision 3 — Questions and Answers
Question 1: What is the role of the Region Proposal Network (RPN) in Faster R-CNN?
- Classifying detected objects into categories
- Generating candidate object bounding boxes (Correct answer)
- Applying non-maximum suppression
- Extracting feature maps from the image
Correct answer: Generating candidate object bounding boxes
The RPN scans the feature map to propose regions that likely contain objects, replacing slower external proposal methods.
Question 2: Which metric is used to measure the overlap between a predicted bounding box and the ground truth box?
- F1 score
- Intersection over Union (IoU) (Correct answer)
- Mean Average Precision (mAP)
- Pixel accuracy
Correct answer: Intersection over Union (IoU)
IoU computes the ratio of the intersection area to the union area of two bounding boxes, quantifying localization accuracy.
Question 3: In image processing, what does a Sobel filter detect?
- Texture patterns
- Color gradients
- Edges via intensity gradients (Correct answer)
- Noise in uniform regions
Correct answer: Edges via intensity gradients
The Sobel filter computes image intensity gradients in horizontal and vertical directions, highlighting edges in the image.
Question 4: What is instance segmentation, and how does it differ from semantic segmentation?
- It assigns one label to the whole image, while semantic segmentation labels pixels
- It labels each pixel with a class, while semantic segmentation uses bounding boxes
- It distinguishes individual object instances with separate masks, while semantic segmentation does not differentiate instances (Correct answer)
- It operates only on foreground objects, while semantic segmentation labels everything
Correct answer: It distinguishes individual object instances with separate masks, while semantic segmentation does not differentiate instances
Instance segmentation identifies and masks each individual object separately, whereas semantic segmentation groups all pixels of the same class together without distinguishing instances.
Question 5: Which technique is used in GANs to generate photorealistic images by training a generator against a discriminator?
- Variational inference
- Adversarial training (Correct answer)
- Contrastive learning
- Triplet loss
Correct answer: Adversarial training
Adversarial training pits the generator against the discriminator in a minimax game, driving the generator to produce increasingly realistic outputs.
Question 6: What problem does batch normalization solve in deep convolutional networks?
- Reduces the number of parameters
- Prevents internal covariate shift by normalizing layer inputs (Correct answer)
- Increases the receptive field size
- Eliminates the need for activation functions
Correct answer: Prevents internal covariate shift by normalizing layer inputs
Batch normalization normalizes activations within each mini-batch, stabilizing training and allowing higher learning rates.
Question 7: In object detection, what is non-maximum suppression (NMS) used for?
- Reducing the learning rate during training
- Removing duplicate overlapping bounding boxes for the same object (Correct answer)
- Normalizing pixel intensities before inference
- Selecting the best anchor box scale
Correct answer: Removing duplicate overlapping bounding boxes for the same object
NMS suppresses all but the highest-confidence bounding box among overlapping detections of the same object, cleaning up redundant proposals.
Machine Learning Computer Vision 3
What is the role of the Region Proposal Network (RPN) in Faster R-CNN?