A Comparison of Equivariant Vision Models with ImageNet Pre-Training

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.

PDF NeurIPSW OpenReview Semantic Scholar

Cite

Text

Klee et al. "A Comparison of Equivariant Vision Models with ImageNet Pre-Training." NeurIPS 2023 Workshops: NeurReps, 2023.

Markdown

[Klee et al. "A Comparison of Equivariant Vision Models with ImageNet Pre-Training." NeurIPS 2023 Workshops: NeurReps, 2023.](https://mlanthology.org/neuripsw/2023/klee2023neuripsw-comparison/)

BibTeX

@inproceedings{klee2023neuripsw-comparison,
  title     = {{A Comparison of Equivariant Vision Models with ImageNet Pre-Training}},
  author    = {Klee, David and Park, Jung Yeon and Platt, Robert and Walters, Robin},
  booktitle = {NeurIPS 2023 Workshops: NeurReps},
  year      = {2023},
  url       = {https://mlanthology.org/neuripsw/2023/klee2023neuripsw-comparison/}
}