Vanessa Wirth

Type: Master Thesis

Description: Real-time 3D reconstruction of dynamic objects is a research topic that has gained wide acceptance in the computer vision community. Among other things, one particular research field of interest is the reconstruction of human motions. In this research field, a large variety of approaches exists, which have shown remarkable results in terms of the reconstruction quality. However, a remaining challenge is the tracking of very fast motions, such as rapid hand and foot movements. In those scenarios, the tracking usually fails since the body parts of a moving human can not be relocated in the succeeding camera frame. Thus, a significant part of correspondence pairs that could be used to track the human motion is lost.
To tackle this problem, the goal of this master thesis is to learn additional correspondences by using a machine learning approach.

The project involves:

1. The generation of an appropriate dataset
2. Training a model that predicts correspondences
3. Using the predicted correspondences in an optimization procedure to solve for human motion parameters of a statistical shape model

Prerequisites:

• Knowledge in at least two of the fields: Computer Graphics, Computer Vision, Machine Learning
• Experience with Python3
• Experience with PyTorch (or willingness to learn)
• Willingness to learn and read source code and literature

Type: Bachelor Thesis / Master Thesis

Description: The tracking of hand movements is a long-standing goal in computer vision with many applications such as in virtual/augmented reality and human-computer interaction. In contrast, tracking has found little traction in the medical context, although its research could sustainably improve the diagnosis of visible diseases, e.g. mobility impairments. In this work, hand motion tracking will therefore be applied to a very specific medical use case in rheumatism research. This use case is about determining the stiffness of the hand based on a grasping motion to pick up an object. A depth camera records this grasping movement and serves as a data basis to estimate the hand movement afterwards.

The goal of this work should be to determine the motion of the hand as robustly as possible across subsequent depth images. An integral part of achieving this goal is the use and comparison of robust cost functions when optimizing hand motions. Other possible challenges to overcome include hand-object interactions, rapid hand motion, and partially occluded hand regions.

Prerequisites:

• Knowledge in either Computer Vision or Computer Graphics (Geometry Processing is a great Plus!)
• Experience with C++
• Experience with Optimization or willingness to learn
• Willingness to learn and read source code and literature