Understanding the Influence of Rendering Parameters in Synthetic Datasets for Neural Semantic Segmentation Tasks

Abstract

Deep neural networks are well known for demanding large amounts of training data, motivating the appearance of multiple synthetic datasets covering multiple domains. However, synthetic datasets have not yet outperformed real data for autonomous driving applications, particularly for semantic segmentation tasks. Thus, a deeper comprehension about how the parameters involved in synthetic data generation could help in creating better synthetic datasets. This work provides a summary review of prior research covering how image noise, camera noise and rendering photorealism could affect learning tasks. Furthermore, we presents novel experiments aimed at advancing our understanding around generating synthetic data for autonomous driving neural networks aimed at semantic segmentation