Maximizing Mutual Shape Information
Md Amirul Islam (Ryerson University),* Matthew Kowal (York University), Patrick Esser (Heidelberg University), Bjorn Ommer (University of Munich), Konstantinos G Derpanis (York University), Neil Bruce (University of Guelph)
The 33rd British Machine Vision Conference
Abstract
We propose a novel training loss for increasing a neural networks ability to encode shape information for object recognition. Our motivation is to improve the generalization ability of recognition networks and their robustness to adversarial attacks and corruptions. We first show that this objective function is differentiable with respect to the network weights, and then we train the network to maximize this objective. We demonstrate the benefits which arise out of models which make decisions based on global object shape rather than local textures. We first show that compared with similar approaches, our method induces a stronger inductive bias in the network towards encoding shape. Furthermore, we also show our model is more robust to adversarial attacks and to distorted images, and can generalize better to out-of-distribution examples. Interestingly, we obtain all these benefits without sacrificing overall performance on ImageNet and transfer learning on various downstream tasks (e.g., semantic segmentation, texture recognition).Video
Citation
@inproceedings{Islam_2022_BMVC,
author = {Md Amirul Islam and Matthew Kowal and Patrick Esser and Bjorn Ommer and Konstantinos G Derpanis and Neil Bruce},
title = {Maximizing Mutual Shape Information},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0909.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by