Enhancement of SSD by Fusing Feature Maps in Multiple Directions

Abstract

With the introduction of the YOLO model, one-stage object detection methods have begun to develop. SSD is one of the most representative work. SSD creatively proposes to use anchor boxes to make predictions on multi-scale feature maps, so that the detection accuracy and speed have reached a high level. In addition, the depth-wise separable convolution has received widespread attention in the recent period of time. Compared with traditional convolution, it has a great advantage in reducing the parameters and calculations of detection framework. Based on this advantage, we propose an efficient object detection method. The core idea of this method is to construct an auxiliary module HLF (Horizontal-Longitudinal Fusion) by applying depth-wise separable convolution and embed it in the detection framework based on SSD to form a more efficient object detection method. On the Pascal VOC 2007 test, HLF SSD can achieve 79.2 mAP (mean average precision) at the speed of 105 FPS (frame per second) with the input size 300×300 using a single Nvidia 1080Ti GPU.