Scholarlink

<p>Weakly supervised object detection (WSOD) is considered important for practical applications since it can train detectors using numerous image-category labels from the Internet. However, there is a large performance gap between weakly supervised and fully supervised detection, which severely restricts its application. On the other hand, we find that WSOD holds great potential in some application scenarios. Specifically, in these scenarios, objects from the same class typically appear in groups and exhibit visual similarity, making the cost of manual annotation extremely high but identification relatively straightforward. Inspired by the above analysis, we establish a practical weakly supervised object detection (PWSOD) benchmark. Specifically, we take two scenes as examples and construct datasets composed of 5,407 and 12,840 images, respectively. Moreover, we focus on addressing the primary challenge in both PWSOD and generic WSOD: missing instances. To solve it, we propose feature augmentation (FA) and multi-instance mining (MIM). FA adds random noise to each proposal feature vector to simulate a new sample feature, thereby enlarging the support of the training distribution. MIM discovers as many positive instances as possible based on the confidence distribution and mitigates the adverse effects of low-quality ones. Thanks to these two modules, our approach can effectively discover previously ignored non-salient individuals. Extensive experiments on PWSOD, PASCAL VOC and MS COCO benchmarks demonstrate our method can bring substantial improvement for WSOD. The datasets are available at <a href="https://pan.baidu.com/s/1dzX3DNIRGDWgfePL_6hGWw?pwd=WSOD">https://pan.baidu.com/s/1dzX3DNIRGDWgfePL_6hGWw?pwd=WSOD</a>.</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

Practical weakly supervised object detection: benchmark and method

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Huan Cao,
  • Zengping Chen

摘要

Weakly supervised object detection (WSOD) is considered important for practical applications since it can train detectors using numerous image-category labels from the Internet. However, there is a large performance gap between weakly supervised and fully supervised detection, which severely restricts its application. On the other hand, we find that WSOD holds great potential in some application scenarios. Specifically, in these scenarios, objects from the same class typically appear in groups and exhibit visual similarity, making the cost of manual annotation extremely high but identification relatively straightforward. Inspired by the above analysis, we establish a practical weakly supervised object detection (PWSOD) benchmark. Specifically, we take two scenes as examples and construct datasets composed of 5,407 and 12,840 images, respectively. Moreover, we focus on addressing the primary challenge in both PWSOD and generic WSOD: missing instances. To solve it, we propose feature augmentation (FA) and multi-instance mining (MIM). FA adds random noise to each proposal feature vector to simulate a new sample feature, thereby enlarging the support of the training distribution. MIM discovers as many positive instances as possible based on the confidence distribution and mitigates the adverse effects of low-quality ones. Thanks to these two modules, our approach can effectively discover previously ignored non-salient individuals. Extensive experiments on PWSOD, PASCAL VOC and MS COCO benchmarks demonstrate our method can bring substantial improvement for WSOD. The datasets are available at https://pan.baidu.com/s/1dzX3DNIRGDWgfePL_6hGWw?pwd=WSOD.