SynCity - Synthetic data for deep learning in automotive applications
| Reference number | |
| Coordinator | Linköpings universitet - Department of Science and Technology |
| Funding from Vinnova | SEK 500 000 |
| Project duration | October 2017 - June 2018 |
| Status | Completed |
| End-of-project report | 2017-03079.pdf(pdf, 404 kB) (In Swedish) |
Purpose and goal
This project has to developed and evaluated new methods for generation of synthetic training data for deep neural networks for applications in computer vision for autonous vehicles, and make publicly available a state-of-the-art open source synthetic dataset for training and evaluation of deep neural nets. The project has investigated which aspects of synthetic data that affects the performance and results when training a deep neural net and strategies for minimising the computational complexity inherent to high accuracy image synthesis and sensor simulation.
Expected results and effects
The data set which has been developed within the project consists of simulated photo-realistic images (simualting a vehicle´s image sensor(s)), as well as the corresponding ground thruth annotations required for training of deep neural networks for semantic segmentation and object detection, which are core components for autonomous vehicles. The evalautions and analysis of the data have generated new insights into how data can be synthesized efficiently and how synthetic data should be used during training and evaluation of an architecture to achieve high performing systems.
Planned approach and implementation
The project is driven by researchers at Linköping University and 7DLabs Inc. (USA) and has through dicsussion with our industrial partners analysed which computer vision applications are most critical for autonomous vehicles, and which aspects of the synthetic data are most central to these. We have, in an iterative process, improved and developed new tools for automatic generation of virtual worlds, synthetic data, and annotations. The evaluations show that synthetic data is an enabling factor in the development of new machine learning algorithms for autonomous vehicles.