Related Experiment Videos
Cross-Attention-Driven Propose-and-Select Generative Data Augmentation for Few-Shot Image Classification
Ying Liu1,2, Liaomo Zheng1,3, Shiyu Wang1,3
1Shenyang Institute of Computing Technology, Chinese Academy of Sciences, Shenyang 110168, China.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a novel Propose-and-Select framework for controllable data augmentation using diffusion models. It enhances few-shot image classification by improving synthetic data quality and diversity without extra training costs.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Generative data augmentation with diffusion models shows promise for few-shot image classification.
- Existing methods face challenges with uncontrollable quality, semantic consistency, diversity, and redundancy.
Purpose of the Study:
- To propose a controllable data augmentation framework for few-shot image classification.
- To address limitations of current diffusion-based augmentation methods.
Main Methods:
- A two-stage Propose-and-Select framework for offline synthetic data curation.
- Utilizing CLIP's zero-shot knowledge for pseudo-labeling, avoiding human annotations.
- Employing adaptive-temperature listwise ranking distillation and multi-objective consistency regularization.
Main Results:
- Achieved 79.58% accuracy on PASCAL VOC and 80.74% on Oxford 102 Flowers.
- Consistently outperformed existing diffusion-based augmentation baselines under a controlled budget.
- Demonstrated improved quality, diversity, and semantic relevance of synthetic samples.
Conclusions:
- The Propose-and-Select framework effectively enhances synthetic data for few-shot classification.
- This approach improves downstream model performance without additional training overhead.
- The method offers a controllable and efficient solution for generative data augmentation.