3 ms·
For some additional context, when the original NeRF paper (https://arxiv.org/pdf/2003.08934.pdf https://arxiv.org/pdf/2003.08934.pdf) was published 2 years ago,
by ath92 5y ago
For some additional context, when the original NeRF paper (https://arxiv.org/pdf/2003.08934.pdf https://arxiv.org/pdf/2003.08934.pdf) was published 2 years ago, it reportedly took at least 12 hours (depending on hardware used of course) to train on the scene with the bulldozer. This has now been reduced to about 5 seconds (!), with realtime rendering of the result.
- hwers 5y agoThe gigapixel example could be done with fourier features which takes about a few minutes to train (on colab-like resources). Definitely still a huge improvement though (and based on more clever hashing techniques than optimization).