Paper Image

Enhancing image classification with inter-class mixing

Published on:

28 March 2024

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Zhicai Wang,

Longhui Wei,

Tan Wang,

Heyu Chen,

Yanbin Hao,

Xiang Wang,

Xiangnan He,

Qi Tian

Bullets

Key Details

Diff-Mix performs inter-class mixing using fine-tuned diffusion models

It balances faithfulness and diversity of synthetic images

The technique expands datasets by generating counterfactual examples

Empirical results show significant gains in few-shot, conventional, and long-tail classification

Diff-Mix works by translating images between classes while retaining backgrounds

AI generated summary

Enhancing image classification with inter-class mixing

This paper introduces Diff-Mix, an innovative data augmentation technique that performs image mixing between classes using text-to-image diffusion models. Diff-Mix is shown to effectively balance faithfulness and diversity of generated images. When used to expand domain-specific image datasets, Diff-Mix leads to marked performance gains in few-shot, conventional, and long-tail classification scenarios.

Answers from this paper

Comments

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

Sign up to comment on this paper

Sign Up