Synthetic Segmentation Dataset Generator for Mining
Author Information
Author(s): Kilian Wilhelm Johannes, Prinsloo Jaco, Vosloo Jan, Taljaard Stéphan
Primary Institution: Faculty of Engineering, North-West University, Potchefstroom, South Africa
Hypothesis
Can a synthetic dataset generator using a 3D modeling framework and raycaster improve the efficiency of creating labeled datasets for deep-level mining applications?
Conclusion
The proposed synthetic dataset generator effectively produces high-accuracy segmentation models for deep-level mining applications.
Supporting Evidence
- The synthetic dataset generator was validated using a deep-level mining case study.
- Segmentation performance metrics improved over training epochs, indicating effective learning.
- The generated dataset contained labeled images of scenes typically found in deep-level mining environments.
Takeaway
This study created a computer program that makes fake images of mining scenes to help computers learn to recognize objects in those scenes without needing to take real pictures.
Methodology
The study used a 3D modeling framework to create synthetic scenes and a raycaster for automatic labeling of objects in those scenes.
Potential Biases
The reliance on synthetic data may lead to overfitting if the generated scenes do not adequately represent real-world variations.
Limitations
The dataset was unbalanced, with a higher representation of pipes compared to other object classes.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website