Synthetic Generation
Utilizing Unreal Engine and NVIDIA Isaacsim to generate synthetic smoke.
Abstract
Synthetic data plays a crucial role in enhancing limited or challenging datasets. One area with a notable scarcity of publicly available data is smoke detection and monitoring. Smoke introduces a unique challenge for computer vision due to its amorphous shape and inconsistent texture. The lack of comprehensive smoke datasets highlights the need for synthetic data generation to augment existing datasets. However, the generation of synthetic smoke and other amorphous objects, as well as the impact of synthetic data quality and quantity on downstream model performance, remains largely underexplored. In this paper, we investigate the impact of synthetic smoke on a limited real-world dataset by utilizing a novel deep learning model specifically tailored to extract smoke features. We used two synthetic smoke pipelines: 1. lower quality but quick to produce smoke generated with Unreal Engine, and 2. higher quality but slower to generate smoke from NVIDIA Omniverse IsaacSim. Across both pipelines, we found SemiS’s performance peaked when synthetic data constituted approximately 30\% of the initial training data. Further, higher quality data enhanced training accuracy by approximately 5\%, compared to a 2.5\% increase achieved with lower quality data. However, Omniverse was $\sim$12\% slower to generate than Unreal. These results demonstrate the usefulness of developing methodologies that determine the value of synthetic data by analyzing their ability to improve model performance in smoke detection and similar applications. While our analysis focuses on smoke plumes, we believe this work establishes a foundation to understand how synthetic data quality impacts the accurate modeling of similar amorphous phenomena.
