A comparative study of object proposals re-ranking methods for object detection
Shuhan Chen, Jindong Li, Xuelong Hu, Ping Zhou · 2016
Object detection is one of the most essential problems in computer vision and has made great progress in recent years, which is mainly contributed by powerful CNN and accurate object proposals. However, most of the existing proposal generation methods suffer from strong localization bias, and achieve high recall by outputting large amount bounding boxes (e.g. 2000). Thus, an effective proposal re-ranking approach becomes crucial to obtain high quality object proposals with few bounding box numbers. In this paper, we make a comparative study of the existing unsupervised objectness measuring approaches to testify their effectiveness and generalization abilities. Experiments on PASCAL VOC 2007 dataset demonstrates that contour is not an adequate objectness cue to pop out high quality proposals, because an object usually has a closed contour, while a good proposal (with high IoU) may not, and considering more objectness cues can get better performance. Thus, more effective objectness cues should be explored and combined together for proposals re-ranking, such more accurate saliency maps, which can greatly benefit subsequent object detection task and is also our future work.