입력 텍스트 프롬프트와 출력 이미지와의 연관성 분석을 이용한 데이터 레이블링 및 테스트베드

초록

In this paper, we aim to validate the efficacy of a dataset created through visualizing the correlation between input text prompts and output images using a text-based generative model. Based on this visualization, we analyze the correlation between the text prompts and the associated output images, and proceed with appropriate labeling to generate an augmented dataset. Various tasks are used to evaluate whether the set goals are achieved, in order to verify the validity of the generated augmented dataset. We precisely analyze the correlation between the input text prompts and the output images, aiming to enhance the inherent accuracy of the data and the performance of the model through this analysis. We demonstrate that the augmented dataset is an effective method to improve the model’s performance, and that automatic labeling and data augmentation methods are also valid.

키워드