Core Concepts
Efficiently learn numerous scenes using a 3D-aware latent space and Tri-Plane scene representations.
Abstract
The article introduces a method to scale Neural Radiance Fields (NeRFs) for learning multiple semantically-similar scenes efficiently. By leveraging a 3D-aware latent space and Tri-Plane scene representations, the training time and memory costs per scene are significantly reduced. The approach involves compressing scene representations to focus on essential information, sharing common information across scenes, and regularizing the latent space with geometric constraints. This results in improved optimization, rendering times, and reduced memory footprint. The method enables learning 1000 scenes with an 86% reduction in training time and a 44% reduction in memory usage.
Stats
Our method reduces effective per-scene memory costs by 44% and per-scene time costs by 86% when training 1000 scenes.
Quotes
"Our method reduces effective per-scene memory costs by 44% and per-scene time costs by 86% when training 1000 scenes." - Article