Master Thesis - Multi-sensor active learning for deep neural networks
Master Thesis & Research Opportunities - Gothenburg, Sweden
Deep neural networks are the state-of-the art method for a number of tasks that are essential for autonomous vehicles, such as object detection and semantic segmentation. However, these networks must be trained on a very large number of human-annotated images, which are costly to obtain. The goal of active learning is to have the neural network inform the selection process, such that the most informative images are selected. As a result, the number of images that must be annotated can be reduced, and we can obtain higher accuracy at a lower cost.
One way to do this is to exploit the fact that we record the same scenes with multiple sensors (namely cameras and LiDAR). The failure modes of detection models trained on each modality are quite different, and we should be able to select which data to annotate based on their disagreement. For example, if an object is confidently detected in the LiDAR scan but not in the camera image, we know that annotating this object will be informative for one of the sensors.
In this master thesis project, you will focus on:
- Developing an active learning method based on combining predictions from camera and LiDAR data
- Applying this method on open-source datasets as well as data recorded by Zenuity's data collection fleet
- Demonstrating that these methods improve the performance of our trained neural networks
We are looking for two students, preferably with good knowledge of:
- Python programming
- Deep learning
- Statistical and probabilistic modelling
Final application date: 2019-11-15. But please submit your application as soon as possible, as we will be screening candidates continuously.
Please send in individual applications with CV, motivational letter and grade transcripts. If you wish to partner with someone, simply note that in your application.
Planned start: 2020-02-01, with some flexibility.
Duration: 30 ECTS
For questions regarding the project, please contact: firstname.lastname@example.org