Analysis of Impact of Synthetic Image Data with Multiple Randomization Strategies on Object Detection Performance
L. C. Adi, T. M. Cheng · 2022
Real-time object detection is a smart tool in assisting in-field human workers. In order to achieve a better performance out of object detection deep learning models, tremendously larger quantities of data are needed to train the models’ network. However, obtaining data is expensive in terms of cost, extremely time-consuming, and prone to errors in the annotated part which reduces the detection performance. Hence, in recent years, many research has been focusing on developing alternatives to supplement real data with synthetic data. One of the most popular techniques to generate them is called domain randomization. In this research, an extensive analysis of the impact of various domain randomization strategies was conducted to discover the impact on object detection among the strategies to generate synthetic data from the example industrial objects. The strategies of object orientation and rendering randomization were comprehensively inspected. Insights on the usage of better synthetic data generation were given to improve object detection performance in selected aspects. The method is especially efficient when extremely limited real data were available, and the delicate employment of synthetic data has strengthened the object detection performance by up to 37% of improvement.