A Comparison of Equivariant Vision Models with ImageNet Pre-training

David Klee, Jung Yeon Park, Robert Platt, Robin Walters

Keywords: Vision, Equivariance, Pre-training

TL;DR: We provide implementations of equivariant vision models with ImageNet pretrained weights for other researchers to use.

Abstract:

Neural networks pre-trained on large datasets provide
useful embeddings for downstream tasks and allow researchers to iterate
with less compute. For computer vision tasks, ImageNet pre-trained models
can be easily downloaded for fine-tuning.
However, no such pre-trained models are available that are equivariant to image
transformations. In this work, we implement several equivariant versions
of the residual network architecture and publicly release the weights after
training on ImageNet. Additionally, we perform a comparison of enforced vs.
learned equivariance in the largest data regime to date.

Submission Track: Extended Abstract

Submission Number: 25