Synthetic training data bias in instance segmentation algorithms
Lena R. Schreiber, Yannick Edward Tarant, Kai Franke · 2024
Protecting infrastructures, particularly buildings, requires the development of automatic detection and mapping technologies for essential components such as pipes in fire extinguishing systems. Despite machine learning advancements in instance segmentation, acquiring extensive, well-annotated real data remains challenging. This leads to the exploration of synthetic images as an alternative solution. This study addresses the limitations of synthetic training data, which can lack realistic features present in real-world images, resulting in biased models and decreased performance. Leveraging Unreal Engine 5 (UE5), a synthetic dataset resembling realworld data from a specific scene is generated. Creating such realistic worlds is time-consuming, so varying domain randomization levels and preprocessing using image filters are explored. With different training set combination, consisting of various distribution of real, synthetic and augmented data, multiple models are trained based on Mask R-CNN and YOLO. After the training phase, an optimization procedure is applied to each model, enabling a comparative analysis of pipe instance segmentation quality for different algorithms based on the composition of the training set. The findings shed light on the efficacy and potential risks of employing synthetic data for training various instance segmentation models. This study provides valuable insights into mitigating challenges associated with data limitations in the training of state-of-the-art neural networks.