Feature Pyramid Networks Quiz

Answers and explanations appear as you take the quiz.

1. What core problem does a Feature Pyramid Network primarily solve?
   • Detecting objects that appear at very different scales
   • Reducing the number of training images needed
   • Removing duplicate bounding boxes after detection
   • Compressing models for mobile deployment
2. Compared to building an image pyramid (resizing and re-running the network), FPN is mainly advantageous because it:
   • Requires no backbone network at all
   • Reuses features already computed inside the network, saving compute
   • Eliminates the need for any convolutions
   • Only works on grayscale images
3. In FPN, what is the role of the top-down pathway?
   • It discards low-resolution features
   • It converts images to feature vectors for classification only
   • It carries semantically rich, deep features back up to higher resolutions
   • It performs non-maximum suppression
4. How are lateral connections typically combined with the upsampled top-down features?
   • Concatenated channel-wise then halved
   • Multiplied element-wise after a softmax
   • Subtracted to compute a difference map
   • Added element-wise after a 1x1 convolution aligns channels
5. Why is a 3x3 convolution applied to each merged feature map in FPN?
   • To smooth the map and reduce aliasing from upsampling
   • To reduce the spatial resolution by half
   • To classify the image into categories
   • To convert features into bounding-box coordinates directly
6. Which later architecture extended FPN with learnable, weighted bidirectional fusion?
   • AlexNet
   • BiFPN, used in EfficientDet
   • LeNet-5
   • Word2Vec
7. How is a top-down FPN feature map typically resized before lateral fusion?
   • Downsampled to a single pixel
   • Rotated 90 degrees
   • Upsampled by a factor of two using nearest-neighbor interpolation
   • Converted into a text sequence
8. What does an FPN preserve by merging shallow high-resolution features with deeper semantic features?
   • Only the original image colors
   • Only the deepest layer’s low-resolution grid
   • The exact training labels without feature extraction
   • Fine spatial detail together with stronger semantic information