Real-data performance evaluation of composite synthetic IR data

Abstract

Achieving state of the art performance with CNNs (Convolutional Neural Networks) on IR (infrared) detection and classification problems requires significant quantities of labeled training data. Real data in this domain can be both expensive and time-consuming to acquire. Synthetic data generation techniques have made significant gains in efficiency and realism in recent work, and provide an attractive and much cheaper alternative to collecting real data. However, the salient differences between synthetic and real IR data still constitute a “realism gap”, meaning that synthetic data is not as effective for training CNNs as real data. In this work we explore the use of image compositing techniques to combine real and synthetic IR data, improving realism while retaining many of the efficiency benefits of the synthetic data approach. In addition, we demonstrate the importance of controlling the object size distribution (in pixels) of synthetic IR training sets. By evaluating synthetically-trained models on real IR data, we show notable improvement over previous synthetic IR data approaches and suggest guidelines for enhanced performance with future training dataset generation.