The paper addresses the challenge of achieving robustness to spatial transformations in computer vision. Traditional methods involve data augmentation or hard-coding invariances, but these can have limitations. The authors propose treating invariance as a prediction problem, using a normalizing flow to predict a distribution over transformations. This approach allows for instance-specific invariance, with the potential to generalize across classes and adapt to out-of-distribution poses. The method shows accuracy and robustness gains on various datasets.
Publication date: 29 Sep 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2309.16672
