FPSSD7: Real-time Object Detection using 7 Layers of Convolution based on SSD
Yuya Yamashige, Masaki Aono · 2019
In recent years, attention has been paid to developing object detection methods from images, based on deep learning. In particular, toward self-driving cars, it is essential to make the detection as accurately as possible, as quickly as possible, as economically as possible. Here, we focus on SSD: Single Shot MultiBox Detector (SSD300) and attempt to improve it with less GPU memory. We propose a new method called “FPSSD7” (Feature Pyramid SSD7), which has less convolutional layers, keeping semantic information. For this purpose, we develop a method to interpolate semantic information between lower layers with high resolution feature maps and upper layer with low resolution features maps. We conducted comparative experiments with the conventional methods using Udacity Annotated Driving Dataset. As the result, we demonstrate that our proposed method outperform the conventional methods.