NVIDIA Instant-NGP: Ubuntu setup and video-to-NeRF tutorial

Avkash Chauhan walks through building NVIDIA Instant-NGP and training a NeRF scene from a video. Uploaded on June 28, 2022, this tutorial documents an Ubuntu 22.04 environment with kernel 5.18, CUDA Toolkit and nvcc 11.7, and Python 3.9 through Conda. It does not identify an Instant-NGP release version. Treat the installation steps as a record of that setup rather than current requirements.

The environment section follows the speaker's build troubleshooting. He attributes compilation problems to a CUDA and nvcc version mismatch, then demonstrates a successful build after updating his environment. He also installs NVIDIA OptiX, which he describes as optional.

For the dataset, he uses a roughly 100-second, 1920-by-1080 video filmed around a centered object. The conversion script extracts frames at two per second and uses COLMAP to recover camera positions. The demonstrated output contains 201 images and a transforms JSON file. He chooses AABB scale 4 for his Titan Xp and warns that larger settings can cause problems depending on the graphics card.

The training walkthrough covers image-path corrections, scene navigation, cropping unwanted surroundings, camera visualization and mesh generation. It also shows saving and reloading a MSGPACK snapshot of the model state, separately from mesh generation, and rendering controls. Later examples cover volumes and neural image reconstruction; the speaker says his older GPU cannot load the 1.2-gigapixel Tokyo image, so he demonstrates a smaller image instead.