Introducing a New Metric for Improving Trustworthiness in Real Time Object Detection

This preliminary study, investigates the possible benefits of using a novel metric in real-time evaluation of outputs generated by Real-Time Object Detection Algorithms. Our primary goal is to improve the reliability of the detection process through the spatial analysis of the sequence of video frames used as input data for a Convolutional Neural Network (CNN). The method focuses on the analysis of the variations between consecutive output values from the CNN. By leveraging established similarity metrics, we try to identify patterns that signal potential instances of false positive predictions and develop a methodology-agnostic assessment of the CNN’s output quality. The paper concludes with some preliminary computational results that support the efficacy and potential applications of the proposed method.