Abstract
In the filed of object detection, a two-stage approach, which includes both region proposal step and object detector, is widely used due to its high detection precision. And the quality of proposals is crucial to the detection result, which should be optimized. However, little attention is paid to the proposal part. In this paper, we put forward a high quality proposal generation network (PGN) to obtain high quality proposals. Specifically, PGN, made up of cascade anchor refinement modules, uses the output of the current module to train the next. The reason is that in single anchor refinement module the quality of output proposals, estimated by IoU distribution, surpasses the input of it. The quality of final proposals in PGN outperforms the predefined anchors after several cascade operations. Then we propose a novel training scheme to further optimize the network. The proposed training scheme employs increasing Intersection Over Union (IoU) thresholds to distinguish positive/negative examples because the regression is relative to the distribution of them. In this way, the former module with a smaller IoU threshold ensures not to omit the positive anchors and the latter module with a larger IoU threshold can make sure that the module regresses more accurately. In addition, we find that cascade operation leads to misalignment between anchor boxes and convolutional features, and also leads to degradation in precision. In order to tackle the mentioned problem, we propose a feature adjustment module, which can better match the refined proposals and their features. Experimental results on MS COCO dataset show that our PGN achieves 65% AR with ResNet-50 backbone, surpassing the region proposal network(RPN) by 8 points. When integrated into the Faster R-CNN and Cascade R-CNN detector, it can boost the performance by 2.2 and 1.3 points respectively.