Scholarlink

<p>Despite significant advancements in deep generative models, generating high-quality training data for object detection remains a challenging task, primarily due to the complex requirements of precise annotations and diverse scenes. To address this challenge, we propose a novel framework for generating high-quality training data using Large Language Models (LLMs). First, we introduce the Layout Enhancement and Diverse Imagery Synthesis Framework (LE-DIS), which leverages LLMs to create diverse target scenes and systematically constructs synthetic data. Next, we propose a CLIP-based image-layout quality metric (CILQM) to evaluate the global consistency and category alignment of synthetic data, ensuring high-quality outputs. Finally, we employ a mixup-based strategy (SRMix) that integrates synthetic and real data to produce diverse training samples, enhancing the model’s stability and adaptability. Extensive experiments on the COCO benchmark demonstrate that our approach significantly improves the performance of both Transformer-based and CNN-based object detection models, highlighting the potential of deep generative models in synthesizing high-quality datasets for object detection tasks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improving Object Detection Models via LLM-Based Training Data Synthesis

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Shi-Ran Ge,
  • Jie Cao,
  • Ran He

摘要

Despite significant advancements in deep generative models, generating high-quality training data for object detection remains a challenging task, primarily due to the complex requirements of precise annotations and diverse scenes. To address this challenge, we propose a novel framework for generating high-quality training data using Large Language Models (LLMs). First, we introduce the Layout Enhancement and Diverse Imagery Synthesis Framework (LE-DIS), which leverages LLMs to create diverse target scenes and systematically constructs synthetic data. Next, we propose a CLIP-based image-layout quality metric (CILQM) to evaluate the global consistency and category alignment of synthetic data, ensuring high-quality outputs. Finally, we employ a mixup-based strategy (SRMix) that integrates synthetic and real data to produce diverse training samples, enhancing the model’s stability and adaptability. Extensive experiments on the COCO benchmark demonstrate that our approach significantly improves the performance of both Transformer-based and CNN-based object detection models, highlighting the potential of deep generative models in synthesizing high-quality datasets for object detection tasks.