Abstract

Training supervised machine learning models with suitable data can be challenging due to the expensive and time-consuming process of collection and annotation, particularly when publicly accessible datasets are not available. This work emphasizes the use of synthetic data to reduce the effort spent on data collection. The study is based on an analysis of three widely-used benchmark datasets and two synthetic datasets one of which was created for this specific task. The insights derived from the preliminary analysis identify the required characteristics of the synthetic data that can consistently lead to improvement in performance metric for the task. In this work, the roles of synthetic data as a supplement to real data is investigated. Precisely, the impact of synthetic data with varying proportions and similarities to real data on the performance of Multiple Object Tracking neural networks based on the Convolutional and Transformer architectures is analyzed. Experiments demonstrate the superiority of using a combination of simulated and real data, where the samples of synthetic data are many folds of the real, and the variance of low-level features is high but limited by the low variance of high-level features. The findings can be applied to other
machine learning tasks and provided guidelines can help improve model performance.