Scaling Object Detection to the Edge with YOLOv4, TensorFlow Lite

Abstract

Recently, on-device object detection has gained significant attention as it enables real-time visual data processing without the need for a connection to a remote server. However, deploying these models on edge devices poses several challenges such as limited computational resources, power constraints, and real-time performance. These challenges are faced in many of the existing methodologies such as Fast R-CNN, and Single Shot Detector. This experiment aims to investigate the scalability of on-device object detection using YOLOv4, CNNs, and TensorFlow Lite. Recent techniques for addressing these challenges include model compression and quantization and the use of hardware accelerators. The proposed objective of this experiment is to evaluate the performance and efficiency of these techniques in the context of on-device object detection and identify potential areas for improvement. The goal is to provide insights that can help guide the design of more efficient and effective edge-based object detection systems.