Learning Diversified Features for Object Detection via Multi-region Occlusion Example Generating.

Object detection refers to the classification and localization of objects within an image by learning their diversified features. However, the existing detection models are usually sensitive to the important features in some local regions of the object. The existing algorithms cannot learn the diversified features regarding to each region effectively, which limit the performance of the model to a certain range. In this paper, we propose a novel and principle method called Multi-region Occlusion Example Generating (MOEG) to guide the detection model in fully learning the features of the various regions of the object. MOEG can generate completely new occlusion examples and it enables our detection model to learn the features of the remaining regions in the object by blocking the important regions in the proposal. It is a general method to generate occlusion examples and it can be implemented to most mainstream region-based detectors very easily such as Fast-RCNN and Faster-RCNN. Our experimental results indicate a 2.4% mAP boost on VOC2007 dataset and a 4.1% mAP boost on VOC2012 dataset compared to the Fast-RCNN pipeline. And as datasets become larger and more challenge, our method MOEG become more effective as demonstrated by the results on the MS COCO dataset.