Artificial Intelligence Computer Vision 2 — Questions and Answers
Question 1: What is 'transfer learning' typically used for in computer vision applications?
- Uploading model weights to cloud storage
- Starting with weights from a model pre-trained on ImageNet and fine-tuning on a smaller domain-specific dataset (Correct answer)
- Running inference on multiple GPUs simultaneously
- Converting models between PyTorch and TensorFlow
Correct answer: Starting with weights from a model pre-trained on ImageNet and fine-tuning on a smaller domain-specific dataset
Transfer learning leverages rich feature representations learned from large image datasets (e.g., ImageNet) and adapts them to new tasks with less data and compute.
Question 2: What does the Intersection over Union (IoU) metric measure in object detection?
- The ratio of correctly classified pixels to total pixels
- The overlap between predicted and ground-truth bounding boxes divided by their union (Correct answer)
- The accuracy of class label predictions
- The speed of inference in frames per second
Correct answer: The overlap between predicted and ground-truth bounding boxes divided by their union
IoU measures how well a predicted bounding box overlaps the ground-truth box; a higher IoU (closer to 1.0) indicates a more accurate localization.
Question 3: Which technique allows a network to locate the image regions most responsible for a classification decision?
- Batch normalization
- Grad-CAM (Gradient-weighted Class Activation Mapping) (Correct answer)
- Max pooling
- Dropout
Correct answer: Grad-CAM (Gradient-weighted Class Activation Mapping)
Grad-CAM uses gradients of the target class score flowing into the final convolutional layer to produce a heatmap highlighting discriminative image regions.
Question 4: What is 'optical flow' in computer vision?
- The brightness distribution across an image
- The pattern of apparent motion of objects between consecutive video frames (Correct answer)
- The dynamic range of colors in an image
- The frame rate of a video camera
Correct answer: The pattern of apparent motion of objects between consecutive video frames
Optical flow estimates the vector field of pixel motion between frames, used for video analysis, action recognition, and object tracking.
Question 5: What is 'non-maximum suppression' (NMS) used for in object detection?
- Removing low-contrast regions from images
- Eliminating redundant overlapping bounding boxes by keeping only the highest-confidence detection per object (Correct answer)
- Normalizing image pixel values before feeding to the network
- Selecting the optimal anchor box sizes for a dataset
Correct answer: Eliminating redundant overlapping bounding boxes by keeping only the highest-confidence detection per object
NMS suppresses duplicate detections by discarding bounding boxes that overlap significantly with a higher-scoring box, ensuring each object has one final detection.
Question 6: What does a Vision Transformer (ViT) do differently from a CNN when processing images?
- It uses pooling layers instead of convolutional layers
- It splits images into patches and processes them as sequences using self-attention, like a Transformer for text (Correct answer)
- It relies on hand-crafted HOG features instead of learned features
- It processes only grayscale images
Correct answer: It splits images into patches and processes them as sequences using self-attention, like a Transformer for text
ViT divides an image into fixed-size patches, flattens and linearly embeds them, then applies Transformer self-attention over the sequence of patch embeddings.
Artificial Intelligence Computer Vision 2
What is 'transfer learning' typically used for in computer vision applications?