Evaluating Object and Text Detectors under the Binary Classification Scenario: A Review

Abstract

With the explosively increasing volume of hateful speech presented with images on the Internet, it is necessary to detect hateful speech automatically. Due to the intense demand for computation from the hateful meme detection pipeline, it is vital to classify the text and non-text images for accelerating the speed of the multimodal hateful speech system. This study reviews the recent development of object and text detection architectures and categorizes them into one-stage or two-stage detectors to better compare accuracy and efficiency. Additionally, this study proposes two datasets as the benchmarks for the binary classification scenario to evaluate two representative object detectors and two state-of-art text detectors on the customized datasets with two types of texts embedded in images. The results indicate that one-stage detectors may not necessarily achieve higher throughputs than two-stage detectors, and the performance of detectors varies depending on the type of image texts. This thesis can contribute to further evaluation of detectors in binary detection tasks.