MobileNet-Tiny: A Deep Neural Network-Based Real-Time Object Detection for Rasberry Pi

Abstract

In this paper, we present a new neural network architecture, MobileNet-Tiny that can be used to harness the power of GPU based real-time object detection in raspberry-pi and also in devices with the absence of a GPU and limited graphic processing capabilities such as mobile phones, laptops, etc. MobileNet-Tiny trained on COCO dataset running on a non-Gpu laptop dell XPS 13, achieves an accuracy of 19.0 mAP and a speed of 19.4 FPS which is 3 times as fast as MobileNetV2, and when running on a raspberry pi, it achieves a speed of 4.5 FPS which is up to 7 times faster than MobileNetV2. MobileNet-Tiny was modeled to offer a compact, quick, and well-balanced object detection solution to a variety of GPU restricted devices.