Visual Object Detection with Score Refinement

Abstract

Robustness of object detection against hard samples, especially small objects, has long been a critical and difficult problem that hinders development of convolutional object detectors. To address this issue, we propose Progressive Refinement Network to reduce classification ambiguity for scale robust object detection. In PRN, several orders of residuals for the class prediction are regressed from upper level contexts and the residuals are progressively added to the basic prediction stage by stage, yielding multiple refinements. Supervision signal is imposed at each stage and an integration of all stages is performed to obtain the final score. By supervision retaining through the context aggregation procedure, PRN avoids over dependency on higher-level information and enables sufficient learning on the current scale level. The progressive residuals added for refinements adaptively reduce the ambiguity of the class prediction and the final integration of all stages can further stabilize the predicted distribution. PRN achieves 81.3% mAP on the PASCAL VOC 2007 dataset and 31.7% AP (15.6% APS) on MS COCO dataset, which demonstrates the effectiveness and efficiency of the proposed method and its promising capability on scale robustness.