Investigating The Best Pre-Trained Object Detection Model for Flutter Framework

Abstract

Object detection is a machine learning task that can detect objects in an image or video. With the rising demand for object detection features, a solution is needed to make it more accessible. This can be solved by integrating an object detection model into Flutter, a framework that can be compiled and used on popular platforms like iOS and Android. We investigated a total of 13 pre-trained models from PyTorch that will be integrated into Flutter. Through our investigation, we found that the YOLOv5 variants provided the best balance between accuracy and speed while boasting a significantly higher accuracy-to-speed ratio compared to the rest. We also found that quantizing the models can reduce their file size and execution time by up to 55% and 26% respectively while retaining comparable accuracies. However, we were not able to integrate them into flutter due to issues that we encountered.