Enhancing Trustworthiness in Real Time Single Object Detection

Abstract

We propose a method for the dynamic evaluation of the output provided by any Real Time Object Detection Algorithm. This work focuses on single object detection from video streams and the main objective is the enhancement of the process with regard to its so-called trustworthiness based on the spatial consideration of the sequence of video frames that are fed as inputs on a Convolutional Neural Network (CNN). To this end, we propose a method that systematically tests the differences between the consecutive values returned by the employed neural network. The process identifies patterns that flag potential false positive predictions based on classic similarity metrics and evaluates the quality of the CNN results in a methodologically agnostic fashion. An extended computational illustration demonstrates the effectiveness and the potentials of the proposed approach.