A Longitudinal Dense Feature Pyramid Network for Object Detection
作者
Jun Wang,Qiujuan Tong,Chan He
标识
DOI:10.1145/3488933.3488939
摘要
Feature Pyramid Network (FPN) is one of the most popular subnetworks in the current object detection model, which deals with the scale change problem in object detection through feature fusion. In feature fusion, when different layers of feature maps are merged into a new feature map, it will inevitably lead to noise increase and information loss. All of these will make a feature pyramid suboptimal. To address these challenges, this paper proposes a new feature fusion network called Longitudinal Dense BiFPN (LD-BiFPN) based on the bi-directional feature pyramid network (BiFPN). LD-BiFPN supplements the missing feature information by reusing the existing feature maps and improves the accuracy of the detector. The focal point of this work is to explore a more efficient feature pyramid network to improve the accuracy of object detection tasks without increasing a large number of parameters. Experiments have demonstrated that the EfficientDet-D0 detector using LD-BiFPN has an Average Precision (AP) improvement of 0.9 points than our re-implemented EfficientDet-D0 detector on the COCO val-2017 benchmark, without bells and whistles.