Region-Based CNNs Quiz

Answers and explanations appear as you take the quiz.

1. What is the core idea of a region-based CNN?
   • Propose candidate regions, then classify and box objects in each
   • Generate a single label for the whole image
   • Compress the image before processing
   • Convert the image to grayscale first
2. Why was the original R-CNN computationally expensive?
   • It trained without image pixels
   • It ran a CNN separately on each proposed region
   • It required one human label per test image
   • It computed only one shared feature map
3. What did Faster R-CNN introduce to replace Selective Search?
   • A bigger CNN backbone
   • A grayscale conversion step
   • A Region Proposal Network (RPN)
   • A recurrent neural network
4. What does RoI pooling accomplish?
   • It deletes regions with low scores
   • It increases image resolution
   • It merges all regions into one
   • It crops and resizes features from a shared feature map for each region
5. What extra capability does Mask R-CNN add?
   • Pixel-level segmentation masks for each detected object
   • Audio transcription
   • Text translation
   • 3D depth from a single pixel
6. What training objective does Mask R-CNN use for its classification, box, and mask branches?
   • Only a pixelwise mask loss with no classification
   • A multi-task loss that combines classification, bounding-box regression, and mask losses
   • A language-model next-token loss
   • A region-proposal loss with no object classification
7. What alignment problem does RoIAlign address compared with RoI pooling?
   • It replaces the Region Proposal Network with Selective Search
   • It removes all need for a shared feature map
   • It avoids the coordinate quantization that can misalign extracted region features
   • It predicts a class label before processing the image
8. How should two-stage and single-stage detector speed/accuracy be compared?
   • Two-stage models are always more accurate and faster
   • Single-stage models cannot detect multiple objects
   • The architectures have identical pipelines
   • Using a specified benchmark and implementation, because the tradeoff is not universal