|
|
Optimization Techniques in Computer Graphics and Appearance Fabrication
Rittig, Tobias ; Wilkie, Alexander (advisor) ; Stamminger, Marc (referee) ; Lensch, Hendrik (referee)
Tobias Rittig Optimization Techniques in Computer Graphics and Appearance Fabrication Optimization, the process of improving an intermediate solution, has been applied in numerous fields of computer science and beyond. Visual computing has throughout been at the forefront of developing new techniques and applying them to synthesize or analyze visual reality. At the latest, the recent boom of deep learning has given attention to large- scale differentiable computation and the application of gradient-descent optimization. This thesis spans parts of this development in four shared first-author publications (three journal, one conference) and three co-authored journal papers. It shows how optimization algorithms are used in two distinct fields of computer graphics. First, the focus is on the emerging field of Appearance Fabrication using full-color 3D printing. We show the benefits of an iterative optimization loop on the sharpness and color accuracy of translucent printouts. Such a loop, consisting of a forward pre- diction and backward refinement, can be composed of various building blocks developed throughout the thesis. The forward prediction can be driven by accurate Monte Carlo path tracing or an approximate neural rendering solution. The backward refinement can rely on a heuristic or a gradient-descent...
|
|
|
Reprezentační neuronové sítě pro diferencovatelné renderování objemu
Nejedlý, Jaroslav ; Rittig, Tobias (advisor) ; Antolík, Ján (referee)
This thesis investigates the possibility of using a representation neural network as a data structure for an appearance prediction model. We present the representation network as an adaptation of the popular neural radiance field network. The representation network is studied on 2D images as well as volumetric data. This thesis also evaluates the outputs of the appearance prediction network which uses the representation network. Our analysis shows a decent quality of the representation network for 2D images and simple volumetric data. However, the outputs of the appearance prediction network are suboptimal. We conclude that the representation network presented in this thesis should be im- proved and the appearance prediction model should be fine-tuned to the representation network as its input. 1
|
|
|
Feature preserving triangle mesh fairing
Ivičič, Michal ; Bubník, Vojtěch (advisor) ; Rittig, Tobias (referee)
In this thesis, we address the problem of smooth interpolatory subdivision of triangle meshes. This problem is more challenging than general smooth subdivision as the original vertices of the mesh cannot be moved to produce smoother surfaces. There are also not that many existing works that cover this problem. As a solution, we use the minimization of a global fairness energy to compute the positions of newly added vertices of the mesh. This approach results in smoother surfaces than the widely used butterfly subdivision scheme, but at the cost of increased computational complexity and potential volume loss in the resulting mesh. We analyze the sources of volume loss and implement a solution to prevent it. Our algorithm also preserves sharp edges and points on the mesh surface. While our solution performs well for basic objects, it may produce artifacts for more complex ones. As a future work, we suggest ways to improve our algorithm to address the artifact occurrences. 1
|
|
|
Artificial Intelligence for the Card Game Durak
Zarlykov, Azamat ; Dingle, Adam (advisor) ; Rittig, Tobias (referee)
Card games with imperfect information present a unique challenge for many common game-playing algorithms because of their hidden game state. The objective of this thesis is to create a framework for implementing and testing various AI agents in the popular imperfect information card game "Durak" to identify the most effective approach in this environment. This paper presents a theoretical and experimental comparison of agents using various techniques, including rules-based heuristics, minimax search, and Monte Carlo tree search. In our analysis, we found that the Monte Carlo Tree Search agent performed the best among the implemented AI agents, whereas the rule-based heuristic agent and the minimax agent were less effective. 1
|
|
|
Iris segmentation
Ramesh, Vishal ; Šikudová, Elena (advisor) ; Rittig, Tobias (referee)
Accurate iris image segmentation is crucial to a range of proposed medical diagnosis and treatment systems. Previous models have worked well with healthy eye images but do not generalize to diseased eye images. We work with a dataset where many subjects have eye diseases or deformities. We analyse the performance of the U-Net, a deep learning architecture for semantic segmentation. Our model was trained on a hand-annotated dataset and tuned to generalise on unseen images. Our model achieves a pixel accuracy of 0.8913 on the test set with a relatively short training time. 1
|
|
|
Real-time voxel visualization and editing for 3D printing
Kužel, Vojtěch ; Rittig, Tobias (advisor) ; Kondapaneni, Ivo (referee)
In this thesis, we explore detailed voxel scene compression methods and editing thereof with the goal to design an interactive voxel viewer/editor, for e.g. a 3D printing appli- cation. We present state-of-the-art GPU compatible data structures and compare them. On top of the chosen data structure, we build standard editing tools known from 2D, capable of changing voxel color in real-time even on lower end machines. 1
|
|
|
Generating High-Precision Navigation Mesh
Sanchez, Luis ; Gemrot, Jakub (advisor) ; Rittig, Tobias (referee)
Navigation meshes are a widely used method for representing the world geometry in a format that can be used by pathfinding algorithms. Frequently used navigation mesh generation algorithms first discretize the input ge- ometry into a grid of voxels and then reconstruct the mesh out of them. This benefits the simplicity and performance of the algorithm, but comes with drawbacks. If the voxels are too large, the navigation mesh is not precise enough and may even have some pathways missing. If the voxels are too small, creation of the mesh takes too long. In this thesis we propose and implement an algorithm that creates a navigation mesh directly from the input geometry without using an intermediate voxel representation. This allows us to preserve original detail where required and results in a more precise navigation mesh. 1
|
|
|
Efficient light transport simulation of participating media in color 3D printing.
Brečka, Bohuš ; Rittig, Tobias (advisor) ; Nindel, Thomas Klaus (referee)
A Monte Carlo light transport simulation is used in scattering-aware color 3D printing pipeline (Elek et al. [2017], Sumin et al. [2019]) to drive an iterative optimization loop. Its purpose is to find a material arrangement that yields the closest match in terms of surface appearance towards a target. As the light transport prediction takes up about 90% of the time it poses a significant bottleneck towards a practical application of this technology. The dense volumetric textures also require a lot of memory. Explicitly simulating every light interaction is particularly challenging in the setting of 3D printouts due to the heterogeneity, high density and high albedo of the media. In this thesis, we explore existing volumetric rendering techniques (Křivánek et al. [2014], Herholz et al. [2019]) and finally engineer a customized estimator for our setting, improving the performance considerably. Additionally, we investigate various storage solutions for the volumetric data and successfully reduce the memory footprint. All the algorithms are available in the form of Mitsuba renderer plugins.
|
|
|
A Methodical Approach to the Evaluation of Appearance Computations
Hruška, Marcel ; Wilkie, Alexander (advisor) ; Rittig, Tobias (referee)
Various rendering techniques often use different approaches to the same aspects of the image synthesis process, mainly due to their complexity and constant development. Excluding global illumination algorithms, appearance descriptions are key distinguishing factors between the rendering systems. These descriptions might include BRDF models, support for spectral color representation, and even integration of advanced phenomena, such as fluores- cence. Unfortunately, as there are no standardized implementations of these features, their computations might not be completely accurate, which may result in their incorrect representation. This thesis describes an evaluation suite that methodically tests rendering algorithms based on their appearance reproduction capabilities. The core of the suite is a set of scenes that test five specific appearance phenomena - polarization, GGX reflectance, fluorescence, iridescence and the overall spectral accuracy. Each test case scenario contains as few scenes as possible while maximizing the number of covered aspects of the tested feature. For the user's convenience, we wrap the scenes inside an automatic workflow that runs the specified test case scenarios and displays the results. As a correctness metric, we provide manually verified reference images that are considered to...
|
|
|
Procedurally Generated Volumetric Cloudscapes for Unity
Koblížek, Jan ; Kahoun, Martin (advisor) ; Rittig, Tobias (referee)
Title: Procedurally Generated Volumetric Cloudscapes for Unity Author: Jan Koblı́žek Department: Department of Software and Computer Science Education Supervisor: Mgr. Martin Kahoun, Department of Software and Computer Science Edu- cation Abstract: The traditional approach to cloud rendering in computer games is based on static sky- boxes or a set of static textures. Volumetric clouds used to be too computationally expensive, but with advances in GPU performance, they were successfully used in recent gaming titles. This thesis presents an implementation of real-time volumetric clouds for the Unity game engine. Clouds are described by multiple textures (both 3-dimensional and 2-dimensional) and rendered using a ray marching algorithm. The resulting implementation allows three types of low altitude clouds - cumulus, stra- tocumulus and stratus. The user can seamlessly transition between different coverages, times of the day, and animate clouds based of the speed and direction of the wind. Clouds support advanced lighting effects such as casting soft shadows and sun shafts. Keywords: clouds, volumetric raymarching, real-time rendering, Unity (game engine) 1
|
guest ::
