Automatic learning for object detection

Abstract

To alleviate the burden of manual image annotation, we propose an automatic learning method to enable object detection. This method mainly consists of the following three aspects: (1) a novel synthetic data generation strategy, which can automatically generate large-scale synthetic data with bounding-box annotations using only semantic concepts of target categories; (2) self-training paradigm combined with synthetic data generation strategy, which mines more information from the unannotated real data through iterative training to improve the performance of the object detector; (3) a simple and effective pseudo box filtering method, which can purify the quality of pseudo boxes during training. Without using any annotations (i.e., image-level annotations and bounding-box annotations) from the PASCAL VOC dataset, our proposed method can obtain 59.3% and 55.1% mAP on PASCAL VOC 2007 and PASCAL VOC 2012, respectively. We also demonstrate the effectiveness of our method on several datasets, including CUB-200–2011, FGVC Aircraft, Stanford Cars, Bird-Aircraft-Car-Dog, and CBCL StreetScenes.