Detail of the student project

Topic:Využití NeRF metod pro optimalizaci 3D rekonstrukce
Department:Katedra kybernetiky
Supervisor:Ing. Michal Polic
Announce as:Bakalářská práce, Semestrální projekt
Description:The accurate 3D reconstruction from RGB-D images plays an essential role in Augmented Reality (AR), camera localization, and defection checks in the industry and construction environments. Most of the current approaches work well in environments rich in textures. However, there is a limitation of a few centimeters (using HoloLens2) of surface accuracy that may be improved by optimization methods employed in NeRF-based approaches, e.g., Mip-NeRF, DS-NeRF, Reg-NeRF, Ref-NeRF, and Efficient-NeRF. Therefore, the student will focus on using RGB-D images in NeRF-based strategies and the loss function modification for fast conversion to highly accurate dense reconstruction. The work compares the state-of-the-art methods for rendering new views from an existing set of RGB images. Further, the student selects the most promising approaches we may combine with RGB-D images.

Follow these steps:
1) Review the state-of-the-art NeRF-based methods (e.g., Mip-NeRF, DS-NeRF, Reg-NeRF, Ref-NeRF, and Efficient-NeRF) for rendering new views from a set of RGB images.
2) Unify the notation and describe the key ideas of the latest methods in your report.
3) Discuss the selection of the most promising methods for optimization of the 3D reconstruction obtained by RGB-D images from the ToF camera or MultiView Stereo (MVS).
4) Run the most promising codes (at least two), evaluate their convergence speed, and compare the accuracy visually.
Responsible person: Petr Pošík