Improving Anchor Boxes And Object Representations For General Object Detection

In recent years,with the rapid development of deep learning techniques,the field of object detection has made significant progress.The success of most state-of-the-art object detectors is derived from the use of feature pyramid and the carefully designed anchor boxes.Despite impressive,the existing methods for constructing feature pyra-mid blindly integrate multi-scale representations on each feature hierarchy.Further-more,these detectors also suffer from some drawbacks brought by the hand-designed anchors.To mitigate the adverse effects caused thereby,we propose a semi-anchor-free network with enhanced feature pyramid for object detection.Specifically,our princi-pal contributions are as follows:1)To better construct feature pyramid,we propose a novel enhanced feature pyramid generation paradigm,which consists of two modules,i.e.,adaptive feature fusion module(AFFM)and self-enhanced module(SEM).The paradigm adaptively integrates multi-scale representations in a non-linear way mean-while suppresses the redundant semantic information for each pyramid level.Thus,a clean and enhanced feature pyramid could be obtained.To the best of our knowledge,this is the first time to explore obtaining better feature representations by suppressing redundant information.2)An adaptive anchor generator(AAG)is designed to yield fewer but more suitable anchor boxes for each input image by introducing little or none a-priori information,where the input of AAG is only the feature representations of the input image.3)Benefiting from the enhanced feature pyramid,AAG is capable of gen-erating more accurate anchor boxes for each input image.Therefore,a semi-anchor-free network with enhanced feature pyramid for object detection could be constructed by in-tegrating the proposed enhanced feature pyramid generation paradigm and AAG.4)We conduct extensive experiments on PASCAL VOC and MS COCO and the experimental results demonstrate the effectiveness of our method.