Deformable Convolutions Quiz

Answers and explanations appear as you take the quiz.

1. What does a deformable convolution add compared to a standard convolution?
   • A larger fixed kernel size
   • Learned 2D offsets that shift each sampling point
   • A pooling step after every layer
   • Random dropout on the weights
2. How are the sampling offsets in a deformable convolution generated?
   • They are hand-coded by the engineer
   • They are fixed constants from a lookup table
   • They are copied from the previous frame
   • An additional convolution layer predicts them from the input
3. Because predicted offsets are usually fractional, how are pixel values sampled at those positions?
   • Bilinear interpolation
   • Nearest-neighbor rounding only
   • Max pooling
   • They are skipped
4. What did Deformable ConvNets v2 (2018) add over the original?
   • Recurrent connections
   • A new loss function only
   • Per-sample modulation weights that scale each point's contribution
   • Removal of all offsets
5. Why are deformable convolutions especially helpful for irregularly shaped objects?
   • Their receptive field can warp to follow the object's actual contours
   • They increase image resolution
   • They compress the model size
   • They convert color images to grayscale
6. Which modern detection model applies the deformable idea to transformer attention?
   • ResNet-50
   • AlexNet
   • U-Net
   • Deformable DETR
7. For an N-point two-dimensional deformable kernel, how many offset components are predicted at each output location?
   • N: one class probability per point
   • N squared: one displacement for every pair of points
   • 2N: one horizontal and one vertical displacement per point
   • Two total, shared by all N points
8. Where does deformable RoI pooling apply adaptive offsets?
   • Only to the detector's output class names
   • To sampling locations within a proposed object's region features
   • To the ground-truth box coordinates before annotation
   • Only to the image file's compression blocks