Related Experiment Video
Updated: Oct 22, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
Investigating Semantic Augmentation in Virtual Environments for Image Segmentation Using Convolutional Neural
Joshua Ganter1, Simon Löffler1, Ron Metzger1
1Faculty of Digital Media, Furtwangen University, 78120 Furtwangen, Germany.
Journal of Imaging
|August 30, 2021
Summary
Creating synthetic data using virtual domains and semantic augmentations offers a fast alternative to real-world data collection for training neural networks. This study found semantic and conventional augmentation methods yield marginally different results for neural network performance.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Acquiring real-world data for training neural networks is costly and time-intensive.
- Synthetic data generation through domain virtualization offers a viable alternative.
- Virtualization allows for rapid data creation and enhanced augmentations with semantic information.
Purpose of the Study:
- To investigate the efficacy of semantic augmentations in virtual datasets for neural network training.
- To compare the performance of neural networks trained with semantically augmented data versus conventionally augmented data.
- To present a novel virtual dataset with semantic augmentations and automated annotations.
Main Methods:
- Development of a virtual dataset incorporating semantic augmentations.
- Generation of automatic annotations for the virtual dataset.
- Comparative analysis of neural network models trained using semantic versus conventional data augmentation techniques for image data.
Main Results:
- Neural network models trained with semantically augmented data showed performance comparable to those trained with conventional augmentation.
- The difference in results between the two augmentation approaches was found to be marginal.
- Semantic augmentation in virtual domains did not lead to significantly superior outcomes compared to traditional methods.
Conclusions:
- Semantic augmentation within virtual datasets provides a comparable alternative to conventional augmentation for neural network training.
- The marginal performance difference suggests that current semantic augmentation strategies may not offer a substantial advantage over established methods.
- Further research into advanced semantic augmentation techniques could potentially enhance the benefits of synthetic data generation.
