Why is it necessary to add the identity matrix in STN network

Question

Why is it necessary to add the identity matrix in STN network

ShichaoJin opened this issue 6 years ago · 2 comments

Answer 1 · 2020-01-28T23:24:39.000Z

Not the original author just a passer by who had a similar question. I think its an insurance policy against a singular matrix though I could be wrong.

Answer 2 · 2020-07-19T02:02:50.000Z

In my experience, in most case, we just need a lightly delta transform of original features, so, add identity make the output x become a delta transform matrix related to original pose, which means the value of x is all near zero. If you didn't, the value of x will be like [1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0], you see, same values of x are near one, but others are near zero. which is not efficient for network to learn.