Multi-scale Feature Fusion Single Shot Object Detector Based on DenseNet

Abstract

SSD (Single Shot Multibox Detector) is one of advanced object detection methods and apparently can detect objects with high accuracy and fast speed. However, detecting small objects accurately remains a problem full of challenges for SSD. To handle this troublesome problem, our paper introduce a multi-scale feature fusion single shot object detector based on DenseNet (MFSOD), which combine the dense convolutional network (DenseNet) with SSD framework. Firstly, we add additional convolutional layers after backbone network to realize multi-scale feature detection. In addition, the feature fusion module is designed to fuse the multi-scale features from different layers, introducing the contextual information in object detection. Finally, we evaluate the proposed method on PASCAL VOC2007 and MS COCO benchmark datasets. The results indicate that our proposed method achieves 78.9% mAP on PASCAL VOC2007 test and 27.1% mAP on MS COCO test-dev2015 at the speed of 23 FPS. MFSOD algorithm outperforms the conventional SSD in aspects of accuracy, especially for small objects, and satisfies the demand of real-time application.