CAwa-NeRF: Instant Learning of Compression-Aware NeRF Features

Omnia Mahmoud, Théo Ladune, Matthieu Gendrin

Abstract: Modeling 3D scenes by volumetric features is one of the promising directions of neural approximations to improve Neural Radiance Field (NeRF) models. Instant-NGP (INGP) introduced multi-resolution hash encoding from a lookup table of trainable feature grids which enabled learning high-quality neural graphics primitives in a matter of seconds. However, this improvement came at the cost of higher storage size. In this paper, we address this challenge by introducing instant learning of compression-aware NeRF features (CAwa-NeRF), that allows exporting the zip compressed NeRF feature grids at the end of the model training with a negligible extra time overhead without changing neither the storage architecture nor the learning model. Nonetheless, CAwa-NeRF is not limited to INGP but could also be applied to any model. CAwa-NeRF minimizes the rate of the learned features by an a simple mathematical approximation for the features entropy. By means of extensive simulations, CAwa-NeRF achieves significant results on different kinds of static scenes such as single object masked background scenes and real-life scenes captured in our studio. It can compress the features grids size down to 6% of the original size with slight improvement in the model quality or down to 2.4% with a negligible loss in the visual quality.

External IDs:dblp:conf/sds2/MahmoudLG24