Scholarlink

<p>This paper proposes a novel approach to enhancing open-world object detection (OWOD) models by combining artificial intelligence generated content and elastic weight consolidation (EWC). To address the issue of low category richness and mitigate catastrophic forgetting, we first utilize stable diffusion with low-rank adaptation (LoRA) fine-tuning to generate customized detection target datasets. These datasets are then employed to train an improved open-world region-based efficient model, incorporating an EWC module to constrain parameter changes during learning new tasks. Experimental results demonstrate that our approach achieves a mean average precision of 84.7% on the generated datasets, significantly improving category richness while mitigating forgetting of previously learned categories. The proposed method effectively balances learning new categories and retaining memory of old ones, advancing the frontiers of OWOD research.</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

Enhancing open-world object detection with AIGC-generated datasets and elastic weight consolidation

  • 期刊论文

  • 发表日期：

    2025年1月21日

查看全文
  • Wenjin Xue,
  • Guowei Xu,
  • Nan Yang,
  • Jian Liu

摘要

This paper proposes a novel approach to enhancing open-world object detection (OWOD) models by combining artificial intelligence generated content and elastic weight consolidation (EWC). To address the issue of low category richness and mitigate catastrophic forgetting, we first utilize stable diffusion with low-rank adaptation (LoRA) fine-tuning to generate customized detection target datasets. These datasets are then employed to train an improved open-world region-based efficient model, incorporating an EWC module to constrain parameter changes during learning new tasks. Experimental results demonstrate that our approach achieves a mean average precision of 84.7% on the generated datasets, significantly improving category richness while mitigating forgetting of previously learned categories. The proposed method effectively balances learning new categories and retaining memory of old ones, advancing the frontiers of OWOD research.