HIWE: Scene Importance Weighted Encoding For Fast Neural Radiance Field Training
Primary Area: representation learning for computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: NeRFs, Implicit representations, Outdoor scene reconstruction, Fast Training
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Neural radiance fields (NeRFs) have emerged as a powerful scene representa-
tion technique to implicitly encode radiance information in space. Recent works
demonstrated that using a grid-based positional encoding to encode 3D radiance
information in space achieves fast training speeds, often requiring only a few min-
utes of training on small-scale synthetic datasets. However, training a NeRF model
that uses a grid encoding on large outdoor scenes requires several hours of train-
ing. In many scenarios, large scenes may have different amounts of detailing at
different regions, with reconstruction/representation quality more important for
some detailing compared to others. Different regions of the scene are however
given equal importance and thus typically no regions of the scene are prioritized
in allocating parameters in the learned model. In this work, we propose a new
grid-based positional encoding technique that integrates scene importance infor-
mation in large scenes to accelerate training. Our encoding flexibly allocates more
model parameters to learn the radiance information in regions of the scene that
are deemed more important. This ensures that the more detailed scene regions are
represented with a larger number of parameters, allowing more detailed radiance
information to be encoded. With our approach, we demonstrate higher quality
representation for the important parts of the scene compared to state-of-art tech-
niques for instant NeRF training, while enabling on-par or faster training times as
state-of-art NeRF models and small model sizes.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7673
Loading