Analysis of Trade-Offs Between Accuracy And Speed of Real-Time Object Detectors for the Tasks of Explosive Ordnance Detection

Abstract

This paper investigates the performance trade-offs between accuracy and inference time in real-time object detection models, particularly for detecting landmines and unexploded ordnance (UXO). Using subsets from the COCO validation dataset, the models' performance is evaluated across various input resolutions. Object detection has seen substantial advancements over the past few decades, propelled by deep learning and enhanced computational resources. However, the balance between accuracy and inference time differs among models and configurations, requiring detailed evaluation to identify the optimal solution for specific applications like landmine and UXO detection. This balance is critical for autonomous robotic systems, where swift and accurate decision-making is imperative.