Podrobnosti studentského projektu

Téma:Neural 3D Reconstruction and Rendering at Scale
Katedra:Katedra kybernetiky
Vedoucí:Torsten Sattler, Dr. rer. nat.
Vypsáno jako:Diplomová práce, Bakalářská práce, Semestrální projekt
Popis:Neural implicit scene representations have recently been shown to be able to model scenes at high levels of detail, allowing accurate 3D reconstructions as well as realistic renderings of the scene from novel viewpoints. Yet, existing approaches are focused on smaller and constrained scenes. The goal of this project is to scale such techniques to larger environments, with images taken at different points in time and at different illumination conditions.

The concrete goals of the project are:
* The student will familiarize themselves with the topic of neural implicit representations for 3D reconstruction and rendering.
* Based on existing frameworks, the student will design and develop an approach based on neural implicit scene representations that can be used for larger and more diverse scenes.
* Using the developed approach, the student will experiment on existing datasets to evaluate the trade-off between being able to represent larger scenes in detail and the resulting run-time and memory requirements.
* https://github.com/NVIDIAGameWorks/kaolin-wisp
* https://github.com/NVlabs/instant-ngp

Neural fields introduction:
* https://neuralfields.cs.brown.edu/cvpr22

Talk on NeRFs:
* https://www.youtube.com/watch?v=HfJpQCBTqZs

* https://waymo.com/research/block-nerf/
* https://niujinshuchong.github.io/monosdf/
* https://zju3dv.github.io/neuralrecon-w/ (https://github.com/zju3dv/NeuralRecon-W)
Za obsah zodpovídá: Petr Pošík