AIFB 01-2022 Bachelor or Master Thesis: Deep Learning Anomaly Detection with Model Contradictions for Autonomous Driving
Institut für Angewandte Informatik und Formale Beschreibungsverfahren (AIFB)
We don’t have autonomous vehicles around us yet because they're not very good at dealing with the many, crazy things we see on the roads of the world. Therefore, anomaly detection is particularly important to detect the unknown and deal with it. In this thesis you will work with deep learning based methods for the detection of anomalies in camera and lidar data based on model contradictions.
- You will evaluate whether model contradictions are suitable to detect true anomalies as well as model failures
- Based on previous work, you will execute and improve a set of three experiments:
- Unknown Lidar Cluster vs CLIP Camera Classifications
- Supervised Lidar Segmentation vs Self-Supervised Lidar Flow Estimation
- Uncertain Semantic Segmentation vs Pseudo-Lidar and Lidar Cluster
- You will thoroughly evaluate the experiments on the multimodal CODA anomaly dataset
as soon as possible
- You study Computer Science or a related course of study
- You are deeply interested in topics such as Autonomous Driving, Robotics, Deep Learning or Computer Vision
- You are able to read and write scientific texts in English
- You are fluid in Python, first experiences with PyTorch
- You show an above-average degree of initiative and commitment as well as a thorough way of working
Das bieten wir Ihnen
- You get exciting insights into our research and gain valuable practical experience
- We use the latest hardware and software. Together with us you work in first-class laboratories (on-site or remotely)
- Regular and extensive support: Weekly 1:1 meetings, bi-weekly student group meetings, monthly 1:1 strategy meetings
- Collaboration with other students to get tips, learn together, and fix issues quickly High-quality theses will be published on KITopen, with the code on GitHub
- We aim to publish this work in an IEEE journal with shared first authorship
Contact person in line-management
For further information, please contact Daniel Bogdoll, email: email@example.com.
Shoot me an e-mail at firstname.lastname@example.org with your CV, grades, and a few sentences why you are interested. No cover letter necessary.
Recognized severely disabled persons will be preferred if they are equally qualified.