Paper Image

Robust image classification with generated datasets

Published on:

5 February 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Hritik Bansal,

Aditya Grover


Key Details

Generated data alone increases classifier robustness but hurts accuracy

Augmenting real data with generated data improves robustness without hurting accuracy

In-the-wild generative models are better than traditional augmentations

More generated data leads to more robustness

Diverse text prompts produce most robust classifiers

AI generated summary

Robust image classification with generated datasets

This paper explores using generated datasets from modern text-to-image models like Stable Diffusion to improve the robustness of image classifiers. The key finding is that augmenting real ImageNet data with equal amounts of generated data leads to models that are more robust to natural distribution shifts like sketches and paintings, without sacrificing accuracy on the original dataset.

Answers from this paper


No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up