Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields

A new technique that combines mip-NeRF 360 and grid-based models such as Instant NGP has been proposed to improve the training of Neural Radiance Fields. This technique uses multisampling to approximate the average NGP feature over a conical frustum, resulting in prefiltered renderings that do not flicker or shimmer, and error rates that are 8%-77% lower than previous techniques while also training 24x faster than mip-NeRF 360. Z-aliasing is addressed in this technique by improving proposal network supervision, resulting in a prefiltered proposal output that preserves the foreground object for all frames in a sequence. The technique’s development was influenced by rendering and signal processing ideas.

To top