SFE-SSD: Shallow Feature Enhancement SSD for Small Object Detection

Abstract:

      SSD (Single Shot Multibox Detector) is one of the best object detection algorithms with both high accuracy and fast speed, but fails to detect very small size object which lacks enough resolution and enough feature information. In order to solve this problem, the majority of existing methods improve accuracy at the cost of a heavy loss of speed. In this paper, we propose SFE-SSD (Shallow Feature Enhancement SSD) to improve performance of SSD model on small object detection based on a novel and lightweight way of feature enhancement module. Firstly, we apply deconvolution on the shallowest feature map in SSD's feature pyramid to enlarge the feature map size and recover more feature details. Then, we introduce semantic information to the enlarged feature map by multi-scale feature fusion. In addition, SFE-SSD is designed to a parallel network structure, which could reduce loss of speed in some degree. Experimental results show that our approach achieved 78.4$\%$mAP and is higher than baseline SSD by 1.2$\%$ on PASCAL VOC2007, especially with significant improvement on small object detection. The testing speed of SFE-SSD is 81 FPS at the cost of a little loss of speed.