Neural Network Convolutional Neural Networks 3

Question 1

What is the primary challenge that fully convolutional networks (FCNs) address?

Accepted Answer

Dense pixel-wise prediction tasks like semantic segmentation

Answer

Classification of very small images

Answer

Training CNNs on video data

Answer

Reducing the number of convolutional layers needed

Question 2

What is dilated (atrous) convolution and what problem does it solve?

Accepted Answer

A convolution with gaps in the filter to expand receptive field without extra parameters or pooling

Answer

A convolution with very small filters for efficiency

Answer

A 3D convolution for volumetric inputs

Answer

A sparse convolution for handling missing data

Question 3

How does instance normalization differ from batch normalization?

Accepted Answer

Instance normalization normalizes per-sample per-channel; batch normalization normalizes across the batch dimension

Answer

Instance normalization normalizes across the batch; batch normalization across spatial dimensions

Answer

They are identical but applied at different layers

Answer

Instance normalization uses learned parameters; batch normalization does not

Question 4

What is non-maximum suppression (NMS) used for in object detection CNNs?

Accepted Answer

Eliminating redundant overlapping bounding box predictions to retain the most confident detections

Answer

Removing neurons with near-zero activations

Answer

Suppressing gradient magnitudes during backpropagation

Answer

Normalizing confidence scores across all detections

Question 5

What metric is used to measure the overlap between predicted and ground-truth bounding boxes in object detection?

Accepted Answer

Intersection over Union (IoU)

Answer

F1 score

Answer

Mean Average Precision only

Answer

Dice coefficient

Question 6

What is the role of anchor boxes in two-stage object detection frameworks like Faster R-CNN?

Accepted Answer

They are predefined reference boxes of various scales/ratios used to propose candidate object regions

Answer

They are the final predicted bounding boxes output by the model

Answer

They anchor the feature maps to specific spatial positions

Answer

They define the receptive field of the RPN head