Object detection based on image pyramid feature fusion and shared detection head

Abstract

In the object detection, the processing of feature fusion and the structure of detection head have an important impact on the detection performance. The current detectors often use the detection pipeline of ‘backbone-feature fusion network-head’. We first propose the Feature Fusion (FF), which constructs a lightweight branching network based on the image pyramid and fuses its extracted features with those of the backbone network, providing a new idea for the focus of feature fusion. In addition, we design the Shared Detector Head (SDH). The main purpose of SDH is to reduce the inconsistency of predictions on feature maps between classification and regression tasks, enhance the interaction between the two, and enhance the detection performance. Our experiments on MS COCO2017 and PASCAL VOC0712 datasets support the above analysis. Based on the above improvements, our approach achieves 0.8% mAP improvements on MS COCO2017. The above experiments prove the effectiveness of our approach.