Reprezentační neuronové sítě pro diferencovatelné renderování objemu
| Název práce v češtině: | Reprezentační neuronové sítě pro diferencovatelné renderování objemu |
|---|---|
| Název v anglickém jazyce: | Neural representations for differentiable volume rendering |
| Klíčová slova: | reprezentační neuronové sítě|neuronové renderování|MLP|NeRF |
| Klíčová slova anglicky: | representation network|neural rendering|MLP|NeRF |
| Akademický rok vypsání: | 2021/2022 |
| Typ práce: | diplomová práce |
| Jazyk práce: | angličtina |
| Ústav: | Katedra softwaru a výuky informatiky (32-KSVI) |
| Vedoucí / školitel: | Tobias Rittig, B.Sc., M.Sc., Ph.D. |
| Řešitel: | skrytý - zadáno vedoucím/školitelem, čeká na schválení garantem |
| Datum přihlášení: | 25.08.2022 |
| Datum zadání: | 25.08.2022 |
| Datum a čas obhajoby: | 14.02.2024 09:00 |
| Datum odevzdání elektronické podoby: | 11.01.2024 |
| Datum odevzdání tištěné podoby: | 11.01.2024 |
| Datum proběhlé obhajoby: | 14.02.2024 |
| Oponenti: | Mgr. Ján Antolík, Ph.D. |
| Zásady pro vypracování |
| The topic of this thesis lies at the intersection of machine learning and volume rendering, specifically differentiable light transport simulation methods that can be used in optimization contexts.
The thesis should investigate how neural representation networks [Mildenhall et al. 20] can be used instead of regular grids to represent the 4D parameters of a volume optimization. Specifically should the networks represent the spatially-varying parameters of a dense scattering medium (single-scattering albedo & optical density). Connected to an existing differentiable neural volume predictor [Rittig et al. 21] one can build a feedback loop that optimizes a target surface color which is formed by the underlying scattering volume. The neural volume predictor requires a multi-resolution input, which is a technique that is also shown in [Barron et al. 21]. This problem is akin to a problem in color 3D printing, although it is less constrained because it is not limited to a discrete set of base materials. Instead, the material parameters can be chosen continuously. The thesis must contain a connection between a representation network and a neural volume predictor, but does not necessarily have to present the whole optimization pipeline. |
| Seznam odborné literatury |
| Representation networks
[Mildenhall et al. 20] NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis, ECCV 2020 [Barron et al. 21] Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields, ICCV 2021 Neural volume rendering [Rittig et al. 21] Neural Acceleration of Scattering-Aware Color 3D Printing, EG 2021 |