Domain Adaptation for object detection and semantic segmentation in autonomous driving

Author: Sergio Casas Pastor

Supervisor: Antonio López

Presentation time: 9:15 h

Virtual Room: 1.2 | Live presentation URL 

Abstract:

Capturing road-annotated data in real environments for autonomous driving is a process that wastes many resources. Collecting data in a virtual environment is a possible solution to overcome this limitation. However, data are not properly transferred between different domains due to variance in scene, objects and camera conditions. The work presented in this document uses two domain adaptation approaches, Cycle GAN and MUNIT, to adapt unlabeled unpaired images from a source domain to a target domain using CARLA simulator. To evaluate the performance of the adaptation, two autonomous driving sub-tasks are applied to the transformed images: road object detection and semantic segmentation. Results show that using adapted data has a significant performance increase in comparison to source domain data in semantic segmentation. Moreover, the adapted models sometimes segment certain elements of images more precise than the target domain model. Consequently, we obtain a good baseline to use data from the simulation in autonomous driving tasks without the need of obtaining annotated data in a real environment. The next stage of the project would include the use of domain adaptation for an end-to-end autonomous driving model.

9 thoughts on “Domain Adaptation for object detection and semantic segmentation in autonomous driving

Leave a Reply

Your email address will not be published.

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>