Semi-Supervised Deep Learning Techniques applied to Traffic Scene Understanding |
|
Mentor: | Rafael Barea Navarro |
Email: | rafael.barea@uah.es |
Phone: | (+34) 918856574 |
University: | University of Alcalá |
Partner Host Institution: | |
Keywords: | Autonomous driving, Robust vision techniques, Scene understanding, Real-time Semantic segmentation, Semi-supervised Deep-learning |
Semi-Supervised Deep Learning Techniques applied to Traffic Scene Understanding
Semantic Segmentation methods play a key role in today’s Autonomous Driving research, since they provide a global understanding of the traffic scene for upper-level tasks like navigation. However, main research efforts are being put on enlarging deep architectures to achieve marginal accuracy boosts in existing datasets, forgetting that these algorithms must be deployed in a real vehicle with images that weren’t seen during training. On the other hand, achieving robustness in any domain is not an easy task, since deep networks are prone to overfitting even with thousands of training images.
This research line proposes to analyze different techniques to be applied to existing deep networks in order to improve their robustness when deployed in any domain. For now, segmentation networks must learn from labeled data in a supervised way to achieve top accuracy. Datasets like CamVid and Cityscapes have hundreds of images, but even their diversity does not guarantee top performance in any unseen scenario in the real world. Some works have addressed the lack of samples by generating synthetic data (e.g. SYNTHIA, CARLA). However, transferring learned features from the virtual domain to the real one is not an easy task. In this context, where data annotation is extremely time consuming and synthetic data isn’t helping, the deep learning community is shifting their efforts to unsupervised models (e.g. GANs) to avoid this high dependence on annotated data. Our proposal is to analyze semi-supervised methods in this context with the aim of implementing real applications for autonomous vehicles.
Departament: | Electronics |
Research Group: | Robotics and e-Safety Unit |
More Information: | www.robesafe.uah.es https://scholar.google.es/citations?user=IktmiSAAAAAJ&hl=es |
Relevants projects on the area: | SmartElderlyCar project (TRA2015-70501-C2-1-R) funded by Spanish MINECO from 2015 to 2019 |
Relevants publications on the area: | 1.- E. Romera, J.M. Álvarez, L.M. Bergasa, R. Arroyo, “ERFNet: Efficient Residual Factorized ConvNet for Real-time Semantic Segmentation”, in IEEE Transactions on Intelligent Transportation Systems. Vol 19, Issue 1, 263-272, January 2018. 2.- Qi Wang, Luis M. Bergasa, José M. Álvarez, “Guest Editorial Introduction to the Special Issue on Robust and Efficient Vision Techniques for Intelligent Vehicles”, in IEEE Transactions on Intelligent Transportation Systems, Vol 19, Issue 1, 129-130, January 2018. |