|
| |
|
|
Appearance matching and fabrication using differentiable material models
Nindel, Thomas Klaus ; Wilkie, Alexander (advisor) ; Chalmers, Alan (referee) ; Magnor, Marcus (referee)
Computing derivatives of code - with code - is one of the key enablers of the machine learning revolution. In computer graphics, automatic differentiation allows to solve in- verse rendering problems. There, parameters such as an objects reflectance, position, or the scattering- and absorption coefficients of a volume, are recovered from one or several input images. In this work, we consider appearance matching and fabrication problems, that can be cast as instances of inverse rendering problems. While gradient-based opti- mization that is enabled by differentiable programs has the potential to yield very good results, it requires proper handling - differentiable rendering is not a shotgun-type prob- lem solver. We discuss both theoretical concepts and the practical implementation of differentiable rendering algorithms, and show how they connect to different appearance matching problems. 1
|
|
|
An Efficient Load-balancing Image Sampler for Path Tracing
Kubíček, Tomáš ; Wilkie, Alexander (advisor) ; Nindel, Thomas Klaus (referee)
In this thesis, we go through the process of improving a research renderer called ART. We focus primarily on the image sampler. We analyse the shortcomings of the stochastic sampler that was present before this thesis was written. Those being the huge memory consumption, lack of real-time visualization, and a limited communication channel between the renderer and the application that is supposed to control it. We then design two samplers that could serve as a replacement. We choose the one which seems to be better suited for the task. We then run a series of tests which serve to finetune parameters of the new sampler and also serve as a comparison between the new and the old sampler. By using the data from the tests we show that we have reduced the memory foot-print while maintaining the speed at mostly the same level as before while enabling working with greater number of cores. 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.
|
|
|
Procedural generation of pencil drawings
Studna, Martin ; Křivánek, Jaroslav (advisor) ; Nindel, Thomas Klaus (referee)
The aim of the thesis is to implement a procedural method which transfers a natural image into a pencil drawing-like style. Our project is written in C++. It uses libraries like OpenCV for image processing and Eigen for linear algebra computations. Since neural networks are frequently questioned, as to whether or not they are better than procedural methods for artistic style reproduction, this work presents also a detailed comparison of both of these approaches. We have re-implemented a selected method for procedural generation of pencil drawing style, bringing several modifications. We compare results of the method with a recently released code for neural network-based drawing generation. The result of this subjective comparison indicates that neural networks maybe be better suited for the generation of pencil-like hatching texture to reproduce shading. On the other hand, the procedurally generated outlined produced by the implemented approach provide more natural renderings.
|
guest ::
