Why is it necessary to add the identity matrix in the transeform.py

charlesq34 / pointnet

PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation

Other

4.64k stars 1.45k forks source link

Why is it necessary to add the identity matrix in the transeform.py #218

Open ShichaoJin opened 4 years ago

ShichaoJin commented 4 years ago

pmorerio commented 3 years ago

I was asking myself the same. In the supplementary material of the paper they mention that

The output matrix is initialized as an identity matrix

I guess this is a way of initializing the transformation very close to the identity. The transformation is then learned in a "residual" fashion (i.e. the matrix is actually parameterized as 1+M). This is the answer I could provide myself. It would be nice to have a confirmation from the authors.

adosar commented 4 months ago

@pmorerio This is also what I am thinking. I was confused with the term initialized since this matrix is always initialized/added at each step (both at training and inference).