EDF-SSD: An Improved Feature Fused SSD for Object Detection
Tanvir Ahmad, Xiaona Chen, Ali Syed Saqlain, Yinglong Ma · 2021
Single Shot Multibox Detector (SSD), which is considered one of the top prominent algorithms for object detection in terms of speed and accuracy. However, in the feature pyramid, conventional SSD uses every layer individualistically and ignores the background information of the objects by taking only the fine-grained details of the objects into account, which drops its accuracy in certain cases especially in heavy occlusion, objects overlapping and small objects. It is a known fact that if the number of feature maps is increased, the efficiency of a deep network is considered to be increased, but just simply rising the number of feature maps will not bring any improvement. Regarding the lack of feature complementarity in conventional SSD feature layers and its low accuracy in certain circumstances, this paper proposed a multi-scale strategy to combine feature maps of a low-level feature with the deconvolutional layers on conventional SSD to improve the efficiency, for simplicity, we called the proposed method EDF-SSD (Enhanced Deconvolution Feature Fused SSD). The proposed EDF-SSD is evaluated on MS COCO and PASCAL VOC 2007 datasets, the experiment results show that EDF-SSD has achieved higher mAP on both datasets by comparing with that of conventional SSD.