Adaptive multiscale feature for object detection

Abstract

In object detection, multiscale features are necessary to deal with objects with different size. Using Feature Pyramid Network (FPN) as the backbone network is a very popular paradigm in existing object detectors, we call this paradigm FPN+. However, feature fusion of FPN is insufficient to express object of similar size but different appearance due to the unidirectional feature fusion. We motivate and present Adaptive Multiscale Feature (AMF), a new multiscale feature fusion method with bidirectional feature fusion, using to solve the one-direction fusion of FPN. AMF module fuses multiscale features from two aspects: (1) shattering features by the way of CLSM; (2) fusing features by the way of channel-wise attention. The proposed AMF improves the expression ability of multiscale features of the backbone network, and effectively improves the performance of the object detector. The proposed feature fusion method can be added to all object detector, such as the one-stage detector, the two-stage detector, anchor-based detector and anchor-free based detector. Experimental results on the COCO 2014 dataset show that the proposed AMF module performs the popular FPN based detector. Whether anchored-free based detectors or anchored based detectors, performance of detector can be improved through AMF. And the resulting best model can achieve a very competitive mAP on COCO 2014 dataset.