Instant-NGP

Instant-NGP (Instant Neural Graphics Primitives) was proposed by Thomas Muller et al. from Nvidia in 2022. The goal of this work is to significantly accelerate the training and rendering speed of Neural Radiance Fields (NeRF) while maintaining high-quality rendering results. Compared with the original NeRF, the most innovative feature of Instant-NGP lies in the introduction of multi-resolution hash tables: instead of using a single large Multi-Layer Perceptron (MLP) network like the original NeRF, it partitions the 3D space into multi-resolution grids, with each grid vertex storing learnable positional encoding features. A hash table is used for compact storage of these features to avoid video memory waste; when querying the color and opacity of a 3D point p, interpolation is performed on the positional encoding features of grid vertices at different resolutions associated with p to form the multi-scale features of p, which can effectively enhance the detail expression capability.

The github repository of instant-ngp is https://github.com/NVlabs/instant-ngp/

In this experiment, students are required to learn the use of Instant-NGP through the following steps:

1) Capture a set of scene images;

2) Train a scene model using Instant-NGP;

3) Specify a rendering camera path and output the rendered video.

For this assignment, only the rendered video as the following needs to be submitted.