Scholarlink

With advancements in data availability and computational power, AI adoption has surged across science and technology. Vision-based methodologies have expanded, particularly in industrial applications, from assembly lines to human-robot interaction. Synthetic data generation in controlled environments enables dataset creation and mitigates challenges like labor-intensive labeling. While synthetic datasets are crucial for object detection training, the domain gap remains a key challenge. This paper explores a two-phase training strategy: first, pretraining with synthetic data, followed by fine-tuning using Active Learning (AL) with uncertainty sampling. We extensively evaluate this approach using well-established benchmarks and YOLOv11 as the detection framework. Additionally, we introduce an industrial Truck dataset, featuring CAD-generated and 3D-printed components of a Truck and Glue-gun. Our findings show that combining synthetic data with AL significantly reduces real data requirements while achieving superior precision, especially when real data is limited.

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

Synthetic Data and Active Learning for Efficient Object Detection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hooman Tavakoli Ghinani,
  • Nimesh Singh,
  • Tatjana Legler,
  • Achim Wagner,
  • Martin Ruskowski

摘要

With advancements in data availability and computational power, AI adoption has surged across science and technology. Vision-based methodologies have expanded, particularly in industrial applications, from assembly lines to human-robot interaction. Synthetic data generation in controlled environments enables dataset creation and mitigates challenges like labor-intensive labeling. While synthetic datasets are crucial for object detection training, the domain gap remains a key challenge. This paper explores a two-phase training strategy: first, pretraining with synthetic data, followed by fine-tuning using Active Learning (AL) with uncertainty sampling. We extensively evaluate this approach using well-established benchmarks and YOLOv11 as the detection framework. Additionally, we introduce an industrial Truck dataset, featuring CAD-generated and 3D-printed components of a Truck and Glue-gun. Our findings show that combining synthetic data with AL significantly reduces real data requirements while achieving superior precision, especially when real data is limited.