Enhanced Single Shot Multiple Detection for Real-Time Object Detection in Multiple Scenes

Abstract

CNN-based object detection architectures have achieved great performances in recent times using SSD, YOLO, and R-CNN. However, using these algorithms for real-time detection suffer from low FPS and accuracy. In this paper, we enhanced the conventional SSD as research has shown that it has higher FPS and accuracy compared to others making it more suitable for real-time object detection. However, this conventional SSD suffers computational complexity and low accuracy for small objects detection. We proposed an enhanced SSD for real-time object detection to improve the accuracy of conventional SSD and reduce its computational complexity with a higher FPS. Our main contribution is at the level of the multi-scale object detection, where we implemented PIV layers for enhanced localization and detection of objects in the feature layers. Furthermore, we introduced extended dilated convolutions with different dilation operations thereby increasing the receptive field and improved the detection of objects. To demonstrate the effectiveness of our proposed method, we first carried out experiments on PASCAL VOC 2007 and PASCAL VOC 2012 and achieved improved performances in mAP of 82.0 and mAP of 85.6 on PASCAL VOC 2007 and PASCAL VOC 2012 respectively at 63 FPS, with input size of 300x300 for a batch size of 8. Using the same experimental approach, we further demonstrated the versatility of the proposed method on the underwater image dataset where we achieved also improved performance in mAP of 79.1. Our experimental results have shown to be an effective alternative for real-time objection detection to the conventional SSD and other state-of-the-art architectures.