Volume rendering with
radiance fields is a technology that will become mainstream in a few years. At
this time, 30 seconds per frame on a top video card means that the amount of
computation is massive. Neural 3D shape representations enable the development
of scenes from angles that are not supported by training data like Deep Voxels.
Transformers from the deep learning area are a similar mapping technique. The
computational cost becomes a concern if we look at the complex trigonometric
formulas for positional encodings and volume sampling.
#volumerendering #transformers #deeplearning
#neuralradiance
#5input
#neuralshaperepresentations
https://arxiv.org/pdf/2003.08934.pdf
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.