4D-Utilities

The goal of this repository is twofold:

1. Neatly organize multiple, static view video files of a single, dynamic scene into standard dataset structures (DyNeRF, Google Immersive, etc) to allow for easier research and development into novel training approaches of dynamic volumetric scenes.

2. Create an end-to-end training and rendering pipeline for novel view creation on cloud GPU's

Features

[x] Portable development environment

[x] Multi-threaded frame extraction utility for DyNeRF

[x] Colmap pipeline for DyNeRF

[x] CLI

[x] poses_bounds.npy utility for custom dataset

[ ] Cloud deployment for spiral render

[ ] Novel view trajectory template

TODO

[ ] Extend frame extraction to include Google Immersive and others

Resources

• DyNeRF datasets

• Pretrained models

Acknowledgement

• Training and rendering scripts from NVlabs' QUEEN

• poses_bounds.npy script from LLFF

• Colmap workflows from 4D Gaussians