Domain Randomization using Synthetic Data for Grease Bin Semantic Segmentation
Sean Chen (Drake University), Khalid Mohammed (Drake University), Nicholas Haisler (Drake University), Haris Mehuljic (Drake University), Nico Gonnella (Drake University), Md Reza (Drake University), Eric Manley (Drake University)
Training robust visual models often requires large, diverse datasets, but collecting real-world data can be costly and time-consuming. To address this challenge, we explore the use of synthetic data for training segmentation models, focusing on container and liquid volume estimation as a downstream application. We generated a synthetic dataset in Blender with varied textures, lighting, viewpoints, and environments to improve segmentation performance. Using HRNetV2, we fine-tuned a model for synthetic-to-real domain adaptation, demonstrating the potential of synthetic data to supplement limited real-world datasets. While this work focuses on segmentation as a precursor to volume estimation, our approach generalizes to other computer vision tasks requiring synthetic data augmentation. The findings contribute to cost-effective model training, reducing the need for extensive real-world data collection.
