There are some question I would like to ask.
For questions 10, I wonder why fully connected layer is the least common instead of average pooling. I can see fully connected layer in neural network for segmentation like U-net.
For question 8, I don’t quite get what ‘vertex are expressed as camera coordinate onto image plane’. Is it mean that we could ignore the extrinsic factor for position of camera (in world coordinate)?
Hi Tsz,
(i) U-net does not have fully connected layers. There are many reasons why fully-connected layers are not employed, but for example, they are quite heavy with respect to the number of parameters, so they are leading to shallower overfitting networks. so in general, fully connected layers are the least common layers you can find in modern architectures.
(ii) yes
