1. Question
STN is an elegant way to add geometric transformation to CNN network. I have some questions:
1. What is the difference between STN and dataset argument methods?
2. Why and how can geometric transformation be inserted into CNN framework? Since geometric transformation is a method in computer vision, is it possible to combine other computer vision methods with CNN?
3. Can STN benefit object detection task? If so, does STN benefit the first stage (feature extraction) of object detection?