Browsing by Author "Winterhalter, Felix"

Now showing 1 - 2 of 2

Open Access
A feature-level analysis of the adversarial robustness of RAFT
(2023) Winterhalter, Felix
Neural networks are ubiquitous in many areas of computer vision, and have gained popularity in the field of optical flow estimation in recent years. While they manage to outperform classical methods in terms of accuracy, they are also known to be vulnerable to adversarial attacks. Since optical flow estimation is a critical component of many computer vision applications, some of which are safety-critical, such as autonomous driving, it is crucial that the vulnerabilities of the methods used are well understood. Further, a modular robustness understanding of a network’s components can help with identifying and fixing vulnerabilities in other networks as well. In this thesis, we perform a robustness analysis of the components of RAFT, a state-of-the-art network for flow estimation, under two different kinds of attacks: The perturbation-constrained flow attack (PCFA), which adds a global perturbation to the whole input image, and an attack based on realistic snow particles, which are added to a three-dimensional scene and then projected onto the input images. We identify different sets of components vulnerable to each attack, and provide some presumptions about the reasons for their vulnerability.
Open Access
General-relativistic polygon rendering
(2020) Winterhalter, Felix
General relativistic visualization addresses optical effects in the realm of general relativity, such as bending of light around massive objects like black holes and stars, affecting the appearance of objects in their proximity. In this thesis, we propose a method for real-time relativistic visualization of triangle meshes in the proximity of black holes, based on polygon rendering and a precomputed table containing information about the distortion and apparent locations of points to a static observer. The visualization process is based on standard polygon rendering through the OpenGL pipeline, but we apply a subdivision to heavily distorted triangles to achieve more accurate results (approximating the curving of straight lines in proximity of a black hole). We consider a static observer with a freely rotating camera, as well as movable objects (considered ‘quasistatic’, i.e. movement itself does not impact appearance, only its location in space). Some results are compared with renders from the ray tracing software GeoVis and we provide additional example visualizations of different objects and scenarios, as well as benchmarks for both the visualization as well as the precomputation-step of our method.