A methodology to select downsized object detection algorithms for resource-constrained hardware using custom-trained datasets

Abstract

Downsized object detection algorithms have gained relevance with the exploration of edge computing and implementation of these algorithms in small mobile devices like drones or small robots. This has led to an exponential growth of the field with several new algorithms being presented every year. With no time to test them all most benchmark focus on testing the full sized versions and comparing training results. This however, creates a gap in the state of the art since no comparisons of downsized algorithms are being presented, specifically using custom built datasets to train the algorithms and restrained hardware devices to implement them. This work aims to provide the reader with a comprehensive understanding of several metrics obtained not only from training metrics, but also from implementation to have a more complete picture on the behavior of the downsized algorithms (mostly from the YOLO algorithm family), when trained with small datasets, by using a fiber extrusion device with three classes: one that has no defects, one that is very similar looking with small changes and one that has a more immediate tell in the difference, showcasing how good the algorithms tell apart each class using two different size of datasets, while also providing information on training times and different restrained hardware implementation results. Providing results on implementation metrics as well as training metrics.