Automatic Annotations for Autonomous Driving Data
Title: Automatic Annotations for Autonomous Driving Data
DNr: Berzelius-2023-197
Project Type: LiU Berzelius
Principal Investigator: William Ljungbergh <>
Affiliation: Linköpings universitet
Duration: 2023-08-15 – 2024-03-01
Classification: 10207


To make autonomous driving (AD) happen, there is a need for diverse datasets. Such datasets are typically large and thus expensive to annotate. This is especially true of the method is to capture the dynamics of the scene, e.g., being able to reason about what the scene will look like at a future point in time. Moreover, it might be infeasible to reach the required diversity by uniformly sampling the world. This project aims to contribute towards solving the data challenge via automatic annotation and data mining of interesting events. We intend to look at one or more aspects of the task. First, we should like to be able to propagate annotations to neighbouring frames, enabling us to annotate a single frame per video and automatically get pseudo-annotations for the others. Second, we intend to experiment with approaches for mining of interesting events, such as cut-ins or sudden braking. Third, we should like to investigate automatic annotation of traffic lights