SyStudio
The data collection processes can be costly and time-consuming when training any machine learning model. Our versatile and easy-to-use synthetic data generation pipeline tackles this issue by allowing users to generate photorealistic image data with pixel-perfect annotation. The user inputs the 3D model of the object and randomization parameters into the toolkit, and it generates any amount of synthetic data needed.
As the primary author, I coordinated and overseed all aspects of the project.
Our synthetic data generation pipeline for SyStudio consists of a six-stage process that transforms 3D models into deployable computer vision models:
- Model Initialization
- Parameter Randomization
- Rendering the Image and Annotation Extraction
- Dataset Generation
- Custom Training of YOLOv5 Object Detection Model
- Evaluation and Iterative Improvement
The work was presented as a research talk during the 2023 Purdue Spring Undergraduate Research Conference.
Citation: Wang, X., Xu, M., Yu, G. (2023, April 13). Synthetic Data Generation for RoboMaster Armor Plate Detection [Oral Presentation]. 2023 Purdue Spring Undergraduate Research Conference, West Lafayette, IN, United States.