Computer Vision

Which statements about Faster R-CNN are TRUE?

• RPN replaces Selective Search
• RPN shares the backbone with the detector
• RPN runs only on CPU
• RPN uses 9 anchors per spatial location (3 scales × 3 ratios)

Which of the following help with class imbalance in segmentation?

• Dice loss
• Pixel accuracy as the metric
• Focal loss
• Weighted cross-entropy

Bottom-up pose estimation (e.g., OpenPose) advantages:

• Runtime ~ O(1) in number of persons
• Always more accurate than top-down
• Uses PAFs for grouping
• Fails when the detector misses a person

DINO's anti-collapse mechanisms include:

• Centering teacher logits with running mean
• Sharpening teacher with low τ
• Negative samples
• Large output dimension (65,536)

Modern transformer upgrades over the original 2017 Transformer include:

• PostNorm
• RMSNorm
• Sinusoidal absolute PE
• RoPE
• GQA

Which TimeSformer attention factorisations were evaluated?

• Space-only
• Joint space-time
• Divided space-time
• Sparse local space-time