search

tscohen / GrouPy

Group Equivariant Convolutional Neural Networks
http://ta.co.nl
Other
353 stars 85 forks source link

size equivariance #14

Closed shuida closed 5 years ago

shuida commented 5 years ago

Hello Dr.Cohen, your group-conv is a great job to preserve rotatation and translation equivariance. I wonder if there is any work about size equivariance, assuming the feature maps and filters stored in infinite arrays.

tscohen commented 5 years ago

Hi Shuida, There are some works that use filters at multiple scales (try searching for "multi scale convolutional networks"), but I don't think any of them are properly equivariant. Building scale-equivariant networks that are numerically stable and really improve performance is an important and challenging problem to work on.

shuida commented 5 years ago

I know about some work using multi-scale filters, which improving performance by choosing several discrete size. But they seem more like tricks rather than solving the problem in theory. If only the continuous size challenge could be overcome, the group-conv will be more perfect.

tscohen commented 5 years ago

I agree. Let me know if you solve it :)