Comprehensive Feature Enhancement Module for Single-Shot Object Detector

Abstract

Recent one-stage CNN based detectors attract lots of research interests due to the high detection efficiency. However, their detection accuracy usually is not good enough. The major reason is that with only one regression step, one-stage detectors like SSD build features which are not representative enough for both classification and localization. In this paper, we propose a novel module, Comprehensive Feature Enhancement (CFE) module, for largely enhancing the features of one-stage detectors. The effective yet lightweight module could improve the detection accuracy with only increasing little inference time. Moreover, we propose two new one-stage detectors by assembling CFEs into the original SSD: CFE-SSDv1 and CFE-SSDv2. The CFE-SSDv1 is of simple structure with high efficiency while CFE-SSDv2 is more accurate and improves dramatically on detecting small objects especially. We evaluate the proposed CFE-SSDv1 and CFE-SSDv2 on two benchmarks for general object detection: PASCAL VOC07 and MS COCO. Experimental results show that CFE-SSDv2 outperforms state-of-the-art one-stage methods such as DSSD and RefineDet on these two benchmarks. Moreover, additional ablation study demonstrates the effectiveness of the proposed CFE module. We further test the proposed CFE-SSDv2 on UA-DETRAC dataset for vehicle detection and BDD dataset for road object detection, and both get accurate detection results compared with other state-of-the-art methods.