DreamDA: Generative Data Augmentation with Diffusion Models
핵심 개념
Proposing DreamDA framework for diverse image synthesis using diffusion models.
초록
Large-scale, high-quality data acquisition is resource-intensive.
Conventional Data Augmentation (DA) techniques lack diversity.
DreamDA utilizes diffusion models for data generation in classification tasks.
Perturbing reverse diffusion process generates diverse samples adhering to original data distribution.
Self-training paradigm introduced for accurate label generation and classifier training.
Extensive experiments show consistent improvements over baselines across tasks and datasets.
DreamDA
통계
Existing generative DA methods inadequately bridge domain gap between real-world and synthesized images.
Diffusion Models introduced to generate highly photo-realistic synthetic images.
Adding Gaussian noise to U-Net bottleneck layer effective for perturbing reverse diffusion process.
인용구
"Our code will be available at https://github.com/yunxiangfu2001/DreamDA."