Robust Object Detection With Inaccurate Bounding Boxes

• DOI: 10.48550/arxiv.2207.09697
• Publication Date: 2022-01-01
• Authors: Chengxin Liu; Kewei Wang; Hao Lu; Zhiguo Cao; Ziming Zhang
• Keywords: Minimum bounding box; Bounding overwatch; Pascal (unit)

Learning accurate object detectors often requires large-scale training data with precise bounding boxes. However, labeling such is expensive and time-consuming. As the crowd-sourcing process ambiguities of objects may raise noisy box annotations, will suffer from degenerated data. In this work, we aim to address challenge learning robust inaccurate Inspired by fact that localization precision suffers significantly boxes while classification accuracy less affected, propose leveraging as a guidance signal for refining results. Specifically, treating an bag instances, introduce Object-Aware Multiple Instance approach (OA-MIL), featured object-aware instance selection extension. The former aims select instances training, instead directly using annotations. latter focuses on generating high-quality selection. Extensive experiments synthetic datasets (i.e., PASCAL VOC MS-COCO) real wheat head dataset demonstrate effectiveness our OA-MIL. Code available at https://github.com/cxliu0/OA-MIL.