An Approach to Improve SSD through Skip Connection of Multiscale Feature Maps.

Xiaoguo Zhang,Kaixin Zhang,Ye Gao,Fei Ye,Qihan Liu

doi:10.1155/2020/2936920

Xiaoguo Zhang, Kaixin Zhang + Show 3 more

Open Access

https://doi.org/10.1155/2020/2936920

Copy DOI

Abstract

SSD (Single Shot MultiBox Detector) is one of the best object detection algorithms and is able to provide high accurate object detection performance in real time. However, SSD shows relatively poor performance on small object detection because its shallow prediction layer, which is responsible for detecting small objects, lacks enough semantic information. To overcome this problem, SKIPSSD, an improved SSD with a novel skip connection of multiscale feature maps, is proposed in this paper to enhance the semantic information and the details of the prediction layers through skippingly fusing high-level and low-level feature maps. For the detail of the fusion methods, we design two feature fusion modules and multiple fusion strategies to improve the SSD detector's sensitivity and perception ability. Experimental results on the PASCAL VOC2007 test set demonstrate that SKIPSSD significantly improves the detection performance and outperforms lots of state-of-the-art object detectors. With an input size of 300 × 300, SKIPSSD achieves 79.0% mAP (mean average precision) at 38.7 FPS (frame per second) on a single 1080 GPU, 1.8% higher than the mAP of SSD while still keeping the real-time detection speed.

Highlights

SSD (Single Shot MultiBox Detector) is one of the best object detection algorithms and is able to provide high accurate object detection performance in real time
Experimental results on the PASCAL VOC2007 test set demonstrate that SKIPSSD significantly improves the detection performance and outperforms lots of state-of-the-art object detectors
To deal with the problem that SSD shows poor performance on small object detection and to maintain a satisfactory detection speed at the same time, we adopt a novel skip connection of multiscale feature maps to SSD, and the overall architecture is illustrated in Figure 2. e main contributions are summarized as follows: (1) SKIPSSD, an improved SSD with a novel skip connection of multiscale feature maps, is proposed to enhance the semantic information and the details of the prediction layers through

Summary

Introduction

SSD (Single Shot MultiBox Detector) is one of the best object detection algorithms and is able to provide high accurate object detection performance in real time. SSD shows relatively poor performance on small object detection because its shallow prediction layer, which is responsible for detecting small objects, lacks enough semantic information To overcome this problem, SKIPSSD, an improved SSD with a novel skip connection of multiscale feature maps, is proposed in this paper to enhance the semantic information and the details of the prediction layers through skippingly fusing high-level and low-level feature maps. Conv4_3_ff skippingly fusing high-level and low-level features; (2) six multiscale feature maps fusion structures over the SSD network, and two feature fusion modules and multiple fusion strategies are designed to investigate the optimal feature fusion framework; (3) experiments on the PASCAL VOC 2007 test set are conducted to compare the performance of SKIPSSD with other state-of-the-art object detectors

Methods

Results

Conclusion