Description
Explain why is scaling and shifting often applied after batch normalization? (Hint: Check the original batch normalization paper)
- Describe the main changes, e.g., architecture change or new losses, made in Faster R-CNN in comparison to Fast R-CNN.
- Describe the differences between the operation of RoIPool and RoIAlign. Explain why RoIAlign is preferred over RoIPool (Hint: Check the paper K. He, Mask R-CNN, ICCV 2017)
- 1) Explain why the encoder-decoder architecture is widely used in semantic segmentation tasks. 2) Does the plain encoder-decoder architecture have potential drawbacks? If so, how can we fix them?
When we apply consecutive 1-dilated, 2-dilated, 4-dilated and 8-dilated 3×3 convolution, what is the final receptive field?
(A question to challenge you!)
1) Even though dilated convolution improves upon standard convolution, what are the potential hard cases for dilated convolution? 2) How will you further improve upon dilated convolution? (Hint: Is there a more flexible form of convolution?)
