Perbandingan Adam dan SGD Menggunakan Faster RCNN MobileNet pada Deteksi Bahasa Isyarat

Abstract

This research aims to evaluate the performance of two different optimizations, namely Adam's optimization and SGD. The faster RCNN backbone used is the MobileNet v3 backbone, and non-maximum suppression (NMS) is applied to strengthen the results of American Sign Language hand gesture detection. Through analysis, it was found that Adam's optimization obtained mAP accuracy of 88.97% on test data, mAP of 86.05% on validation data, and average f1-score of 88.54% on test data. SGD optimization showed mAP accuracy of 88.35% on test data, mAP of 87.23% on validation data, and an average f1-score of 84.56% on test data. Thus, although both optimizations show similar performance, Adam's optimization has a slightly superior performance in terms of mAP and f1-score on the test data. This research implies that using Backbone MobileNet v3 and Adam's optimization can be a practical approach to support American Sign Language hand gesture detection.