Work

Unsupervised feature learning for road segmentation with deep neural networks using generative models and auxiliary tasks

Year: 2017
Type: dissertation
Abstract: Deep learning techniques have lately shown promising results when applied to computer vision problems.An area benefiting from these techniques is the development of autonomous vehicles.Neural networks can be used to obtain an accurate perception of the vehicle surroundings, but need to be trained with large amounts of data with high quality annotations to achieve high performance.These annotations are generally expensive and time consuming to create.Hence there is a demand for reducing the amount of annotated data necessary for training an accurate model.One way to address this problem is by leveraging the structure present in the data itself, using unsupervised training methods.The main objective of this thesis is to explore such methods applied in the area of autonomous vehicles.The primary focus of this thesis is the KITTI road detection task of performing semantic segmentation of road area.The thesis investigates different semi-supervised methods to perform this task by utilizing large unannotated datasets together with annotated datasets of limited size.Different types of generative models are explored, including generative adversarial networks and variational autoencoders.The generative models are used to perform feature extraction using unannotated data.The explored methods for unsupervised feature learning are quantitatively evaluated by comparison to different baseline methods.It is shown that unsupervised feature learning can be used to improve the performance of models for road segmentation.
Language: English
Open Access status: green
OpenAlex ID: W2732862154
Cites: 24
Cited by:
Related to: 10