Rotation equivariant vector field networks [1] (to appear in ICCV 2017)
Using MatConvNet 1.0-beta25 the intallation can be done by running vl_contrib('RotEqNet')
. For older versions it can also be unziped manually.
Run the setup script setup_RotEqNet.m
from Matlab and then main.m
to download the MNIST-rot dataset, train and test a rotation invariant classifier.
Anders Waldeland, from the Norwegian Computing Center, has been working on a Pytorch remake of RotEqNet! The code is available here.
[1] Marcos, D., Volpi, M., Komodakis, N., & Tuia, D. (2016). Rotation equivariant vector field networks. arXiv preprint arXiv:1612.09346.