Extending Rapid Class Augmentation to a Single-Shot-Detector Object Detection Framework

Abstract

This paper describes how eXtending Rapid Class Augmentation (XRCA) optimization can be integrated into a modern single-shot detector (SSD) architecture to enable fast and efficient progressive learning of new objects. The key distinguishing property of XRCA optimization is the incorporation of memory from previously learned classes into its weight update equations. This allows XRCA models to optimally learn new types of objects using just the new object training data. The new XRCA-SSD object detection framework replaces the traditional SSD's prediction heads with the XRCA prediction heads that use different XRCA optimization modes to update the weights. The mean average precision (mAP) performance metric for a SSD model trained using XRCA versus stochastic gradient descent is compared and the XRCA-SSD trained model is shown to greatly outperform the SGD-SSD model by largely mitigating the impact of catastrophic forgetting during new object augmentation. We expect the new XRCA-SSD framework to be especially relevant for real-time progressive learning applications where rapid training times are critical, and compute and memory are often limited.