Machine Learning Computer Vision 5 — Questions and Answers
Question 1: What is panoptic segmentation?
- A method that only segments foreground objects with instance masks
- A unified task combining semantic segmentation of background stuff and instance segmentation of foreground things (Correct answer)
- Segmentation using panoramic 360-degree images
- A technique for segmenting 3D point clouds
Correct answer: A unified task combining semantic segmentation of background stuff and instance segmentation of foreground things
Panoptic segmentation unifies semantic segmentation (for background regions) and instance segmentation (for countable objects) into a single coherent output.
Question 2: Which attention mechanism in Vision Transformers (ViT) allows the model to relate different patches of an image to each other?
- Convolutional attention
- Channel attention (SE block)
- Self-attention (Correct answer)
- Spatial pyramid pooling
Correct answer: Self-attention
Self-attention computes relationships between all pairs of image patches, enabling ViT to capture long-range dependencies without convolutions.
Question 3: What is image super-resolution in deep learning?
- Classifying images at multiple resolutions simultaneously
- Reconstructing a high-resolution image from a low-resolution input (Correct answer)
- Detecting objects at very small scales
- Generating multiple views of a scene from a single image
Correct answer: Reconstructing a high-resolution image from a low-resolution input
Super-resolution models learn to reconstruct fine details and textures in high-resolution outputs from coarse low-resolution inputs.
Question 4: In stereo vision, what is a disparity map used to compute?
- The color distribution across the scene
- Depth information from two horizontally offset cameras (Correct answer)
- Motion vectors between consecutive frames
- Semantic labels for each pixel
Correct answer: Depth information from two horizontally offset cameras
Disparity (the horizontal shift of corresponding pixels between left and right camera images) is inversely proportional to depth, enabling 3D reconstruction.
Question 5: What does mean Average Precision (mAP) measure in object detection evaluation?
- Average pixel accuracy across all semantic classes
- The average of per-class Average Precision scores across multiple IoU thresholds and classes (Correct answer)
- The mean inference speed across different model sizes
- Average localization error of bounding box predictions
Correct answer: The average of per-class Average Precision scores across multiple IoU thresholds and classes
mAP aggregates the area under the precision-recall curve for each class and averages across all classes and IoU thresholds, providing a comprehensive detection metric.
Question 6: What is the key innovation of the Inception module used in GoogLeNet?
- Using only 1×1 convolutions to reduce computation
- Applying convolutions of multiple kernel sizes in parallel and concatenating the outputs (Correct answer)
- Replacing fully connected layers with global average pooling
- Introducing residual skip connections between layers
Correct answer: Applying convolutions of multiple kernel sizes in parallel and concatenating the outputs
The Inception module applies 1×1, 3×3, and 5×5 convolutions (plus pooling) in parallel, letting the network learn features at multiple scales simultaneously.
Question 7: Which technique generates visual explanations showing which image regions most influenced a CNN's classification decision?
- T-SNE visualization
- Grad-CAM (Gradient-weighted Class Activation Mapping) (Correct answer)
- Principal Component Analysis
- UMAP embedding
Correct answer: Grad-CAM (Gradient-weighted Class Activation Mapping)
Grad-CAM uses gradients of the target class flowing into the final convolutional layer to produce a coarse localization heatmap highlighting important regions.
Machine Learning Computer Vision 5
What is panoptic segmentation?