Synthetic Data Generation: An Evaluation of the Saving Images Pipeline in Unity

In this paper, we propose a new framework, CUDA DirectX 11 Interop, for the Saving Images Pipeline that can exceed current performance standards. Unity generates synthetic data to supplement the large amounts of training data required by neural networks. The Saving Images Pipeline saves synthetic data to disk, but is not optimized due to overhead costs. Current (Linear and Coroutines) and alternative (DirectX 11 and CUDA OpenGL Interop) solutions were evaluated in terms of execution times, image resolutions, and JPEG qualities. The DirectX 11 solution exhibited the best performance as it met the 40 ms baseline. Using CUDA Interoperability with the DirectX 11 solution creates a new framework, CUDA DirectX 11 Interop, that exceeds the performance of all solutions examined.