Photorealistic Neural Appearance Transfer - ETHZ Master thesis at NVIDIA (2021)
Proof of concept work that showed it was possible to transfer the appearance of an object to a NeRF scene through 2D images only and retain parametric control.
- An appearance transfer pipeline, that captures the desired look and retains parametric properties, like lighting.
- A depth oracle network that can learn depth information without ground truth information and efficiently render NeRF scenes (down to 8 samples per ray instead of 64 samples).
- A method to combine a depth oracle and a NeRF network to preserve the shape of a NeRF scene during appearance transfer.
This was a joint project between ETHZ (CGL) and NVIDIA.
Thanks to Jonathan Granskog (NVIDIA), Marios Papas (Disney Research), Fabrice Rousselle (NVIDIA), Jan Novák (NVIDIA) & Hendrik Baatz (NVIDIA & Indoor Astronaut) for helping and advising me through my thesis.