Combining Synthetic Images and Deep Active Learning: Data-Efficient Training of an Industrial Object Detection Model

This paper combines synthetic images generated via physics-based rendering with deep active learning for industrial object detection. The results demonstrate that synthetic images significantly improve model performance, especially when real training data is scarce. The hybrid approach of using synthetic and selectively labeled real images outperforms models trained solely on real data, reducing annotation costs and addressing the sim-to-real domain gap. The study emphasizes the value of synthetic data in data-efficient training for industrial applications. To read the complete article click on the link... https://www.mdpi.com/2313-433X/10/1/16