Vanessa Wirth

Type: Master Thesis

Description: With the rise of low-cost RGB-D devices (Google Tango, Microsoft Kinect, etc.), a rapid movement of research towards 3D scene reconstruction emerged. Whereas direct reconstruction methods are prone to geometric incompleteness and noise, another research direction tackles the reconstruction problem in the synthetic domain and makes use of Computer-Aided-Design (CAD) models, which are aligned to the respective position of the “real-world” objects and, thus, fill in the missing object geometry. One particular task in this research direction is to find the most suitable CAD model with respect to a real-world object, further referred to as CAD model retrieval. Most learning approaches [1][2] for CAD model retrieval operate on 3D input. However, we also saw 2D image-based approaches [3], which are more practical in terms of the required ground-truth data. With respect to the latter, this work aims to develop a learning approach for accurate CAD model retrieval on the basis of image-based data.

Prerequisites:

• Knowledge in at least one of the fields: Computer Graphics, Computer Vision, Pattern Recognition
• Experience with Python3
• Experience with PyTorch (or willingness to learn)
• Basic experience with C++ and OpenGL
• Willingness to learn and read source code and literature

[1] https://arxiv.org/abs/1811.11187
[2] https://arxiv.org/abs/1906.04201
[3] https://arxiv.org/abs/2007.13034

Type: Master Thesis

Description: With the rise of low-cost RGB-D devices (Google Tango, Microsoft Kinect, etc.), a rapid movement of research towards 3D scene reconstruction emerged. Whereas direct reconstruction methods are prone to geometric incompleteness and noise, another research direction tackles the reconstruction problem in the synthetic domain and makes use of Computer-Aided-Design (CAD) models, which are aligned to the respective position of the “real-world” objects and, thus, fill in the missing object geometry. Recent learning approaches in this research direction aim to estimate the optimal CAD model alignment on the basis of video [1] and/or RGB-D [2][3] input. We also saw image-based approaches [4], which are more practical in terms of the required ground-truth data. This work is inspired by the latter and aims to develop a learning approach for CAD model alignment on the basis of image-based RGB data. Contrary to [4], this work uses learned canonical representations and a different optimization method for CAD model alignment.

Prerequisites:

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

[1] https://arxiv.org/abs/2012.04641
[2] https://arxiv.org/abs/1811.11187
[3] https://arxiv.org/abs/1906.04201
[4] https://arxiv.org/abs/2007.13034

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. Besides model or template-based methods, in which a static 3D reconstruction of the object is performed first, another research direction deals with reconstructing the hull of an object during its movements. Many methods are based on the well-known volumetric fusion approach, in which the object is reconstructed gradually from the fusion of several successive depth images. In this process, the fused object is usually stored in the form of an implicit function, which in turn is explicitly encoded by a volumetric grid. However, several approaches use very inefficient data structures to represent this volumetric grid. As a result, the memory footprint of the volume becomes very large and only very coarsely resolved explicit representations of the reconstructed object can ultimately be stored. Therefore, the goal of this master thesis is to adapt volumetric data structures from the rendering domain to the 3D reconstruction of non-rigid objects to further improve the quality of objects.

Prerequisites:

• Knowledge in at least one of the fields: Computer Graphics, Computer Vision
• Experience with C++
• Willingness to learn and read source code and literature