Description

keywords: anomaly detection, deep-learning, point cloud data, 3D.

Recently, two new data sets on the topic of anomaly detection in point cloud data have been released. A point cloud is an unstructured collection of 3D points, which may be derived from 3D scanning an object. The current state-of-the-art methods do not achieve convincingly high detection rates for industrial application. While the TNT has a history of successful anomaly detection methods, there is currently no work in the domain of 3D point clouds. This thesis should look for new approaches to fill this gap. The existing literature on (learning-based) point cloud processing needs to be analysed. From that point on, suitable methods should be chosen (or derived from scratch) to tackle the anomaly detection task. The exact methods, details, and scope of this work can be discussed depending on previous experience and motivation.

• Real3D-AD: A Dataset of Point Cloud Anomaly Detection
• Towards Scalable 3D Anomaly Detection and Localization: A Benchmark via 3D Anomaly Synthesis and A Self-Supervised Learning Network
• Anomaly Detection Papers from the TNT

If you are motivated and interested in this project, please send me a short email with the important information about you (relevant lectures you have attended / programming skills) and why you are interested.

Requirements

• Motivation
• Willingness to work independently
• Knowledge of these topics may be beneficial:
• Machine Learning
• Deep Learning
• Good programming skills required in:
• Python and scientific computing libraries
• Deep-Learning libraries (e.g. PyTorch)

Contact person: Mathis Kruse