Analysis of metadata and machine learning with currently largest dataset (16TB)

Open Human-Computer Interaction Computer Science Master Project Thesis No Affiliation

Motivation

In virtual reality, it is possible to identify people by their motion data with machine learning (Miller R et al. (2022), Rack et al. (2022)). For this, it is necessary to train the neural network with properly prepared datasets, because poorly prepared data can lead to sub-optimal learning results (Rack et al. 2023). Last year, Vivec et al. (2023) published the largest dataset to date with more than 100,000 users who play Beat Saber. We have now received a much larger dataset from him that we can use and publish, consisting of 16 TB with over 200,000 users. However, this dataset is currently available in raw form, so careful analysis is essential to make it usable for machine learning. If this dataset is properly prepared, it will enable us to train and test existing models in novel ways, as it provides access to an unprecedented amount of user data.

Goal

The focus of this project is on analysing and viewing the extensive dataset. The aim is to extract relevant meta-information and analyse the movement files using simple methods in order to identify erroneous data. The data prepared and validated in this way is to be made available in a database that enables quick access to datasets that fulfil specific conditions. The dataset will then be used for initial training with existing neural networks.

Tasks

• Inspection and evaluation of the existing data for completeness and consistency.
• Development of software that extracts the metadata from the dataset and transfers it to a database.
• Training of an existing neural network with a large amount of data from the large dataset.
• Evaluation and presentation of results

Prerequisits

• Experience with Python
• Interest in developing and improving data analysis methods
• Basic knowledge of machine learning

Literature

1. Christian Rack, Vivek Nair, Lukas Schach, Felix Foschum, Marcel Roth, Marc Erich Latoschik, Navigating the Kinematic Maze: Analyzing, Standardizing and Unifying XR Motion Datasets. 2024.
2. Christian Rack, Lukas Schach, Marc Latoschik, Motion Learning Toolbox. A Python library for preprocessing of XR motion tracking data for machine learning applications. 2023.
3. Vivek Nair, Christian Rack, Wenbo Guo, Rui Wang, Shuixian Li, Brandon Huang, Atticus Cull, James F. O’Brien, Marc Latoschik, Louis Rosenberg, Dawn Song, Inferring Private Personal Attributes of Virtual Reality Users from Head and Hand Motion Data. 2023.

Contact Persons at the University Würzburg