DrlNet: Blind object proposal quality assessment with discriminative response learning

Abstract

Object proposal quality assessment without ground truth as reference is a challenging task. Some existing methods measure the quality with hand-crafted metrics for subjective metrics, such as objectness and foreground confidence. Recently, deep learning is adopted for direct assessment for quantifiable metric, such as Intersection over Union (IoU). However, we find that IoU, the commonly used quality metric, is far from fully describing the quality of an object proposal. Proposals with the same IoU score may carry totally different amount of discriminative attribute. We introduce a new metric named Discriminative Information Richness (DIR) to characterize the discriminative degree of the given object proposal. DIR is derived from the response intensity of the projected deep feature maps, whose high correlation response indicates the discriminative regions. Besides, we design a convolutional neural network named DrlNet to simultaneously predict IoU scores and perceive the richness of the identification information. DrlNet is defined as a multi-metric joint deep regression network for both spatial covering prediction and discriminative information richness perception. Compared with the solely IoU based models, DrlNet can provide more comprehensive quality assessment. We perform comprehensive experiments on both PASCAL VOC dataset and COCO dataset. The experimental results show that our DrlNet performs well on both proposal selection and object detection tasks. Particularly, experimental results on COCO dataset demonstrate the good generalization ability of the proposed model.