Real-time Robust Object Detection Using an Adjacent Feature Fusion-based Single Shot Multibox Detector

Abstract

A single shot multibox detector (SSD) is used as a baseline for many object detection networks, since it can provide sufficiently high accuracy in real time. However, it cannot deal with objects of various sizes, because features used in an SSD are not robust to multi-scale objects. To solve this problem, we present an improved feature pyramid for using multi-scale context information. The proposed feature pyramid fuses only adjacent features of the conventional SSD to achieve high accuracy without decreasing the processing speed. Our detector, with a 320×320 input, achieved 79.1% mean average precision (mAP) at 63 frames per second on a Pascal Visual Object Classes Challenge 2007 test set using a single Nvidia 1080 Ti graphics processing unit. This result shows better performance than existing SSDs.