Computer Vision Quiz

Computer Vision Quiz

Check your understanding across the course — anchor box mechanics and NMS, U-Net skip connections, backbone trade-offs, CLIP's shared embedding space, and what augmentations do in self-supervised learning.

Exercise

In anchor-based object detection (Faster R-CNN, YOLO), why do models predict offsets from anchor boxes rather than predicting absolute bounding box coordinates?

Need a hint?

Think about the scale of the values and what makes learning easier.

Exercise

Standard NMS removes boxes with IoU > threshold relative to the kept box. Soft-NMS replaces this hard suppression with score decay. When is Soft-NMS preferred?

Need a hint?

When would hard removal of a box be a mistake?

Exercise

A self-driving car system needs to label every pixel as road/sidewalk/car/pedestrian/sky. Which type of segmentation is this, and which model architecture is appropriate?

Need a hint?

The system needs pixel-level labels but doesn't need to distinguish between individual cars.

Exercise

U-Net was originally designed for biomedical image segmentation where labeled training data is scarce. Which architectural choice makes U-Net effective with limited training data?

Need a hint?

Think about data augmentation strategy and architectural choices that reduce the number of labeled samples needed.

Exercise

You need to deploy an object detector on a smartphone (limited compute, 30 fps required). Which backbone choice is most appropriate and why?

Need a hint?

Consider the compute/accuracy trade-off — not all backbones are created equal.

Exercise

CLIP classifies images into 1,000 labels it was never trained on. What makes this zero-shot classification possible?

Need a hint?

Where do the image embedding and the prompt embeddings live?

Exercise

In contrastive self-supervised learning (SimCLR-style), what role do the data augmentations play?

Need a hint?

The model is trained to give two augmented views of one image the same representation.

Exercise

Self-supervised learning invents labels from unlabeled data (predict a masked patch, match two augmented views). Which slot of the ML project loop does this most fundamentally address?

Need a hint?

It manufactures supervision — it changes what training signal is available.

Exercise

Transfer test — a technique this course did NOT teach: label smoothing, which replaces hard 0/1 targets with soft ones (e.g. 0.9/0.1). Which slot does it modify?

Need a hint?

It changes the training targets — the loss the model minimizes.