Add NeRF--

[mwtarnowski]

NeRF--: Neural Radiance Fields Without Known Camera Parameters
by Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, Victor Adrian Prisacariu.

An end-to-end framework for training NeRF models given only RGB images, without pre-computed camera parameters.

[kwea123]

As I observed from a lot of experiments I have conducted with NeRF (on real world car scenes), it is true that NeRF doesn't require any "correct" intrinsics+extrinsics to achieve good results. By "correct" I mean careful camera calibration using checkerboard and so on. Rather, using these correct values, surprisingly, produces inferior performance. Hence now I'm always using colmap estimated values even if they are not "correct", and it also seems that there exists a lot of possible intrinsics+extrinsics combinations that make NeRF work well.

This work confirms my experience that many calibrations work for NeRF, and it is possible to learn one of them. I'm looking forward to experimenting how it performs with my data and how much longer it takes compared to NeRF+COLMAP (there are more parameters to optimize, so I expect a longer training time, but at the same time it saves COLMAP time).

[yenchenlin]

Hello @mwtarnowski , thanks again!

Same as the previous PR, I have put the citation of this paepr here. Could you update your PR to cite this txt file?

Very helpful and timely PR!

[yenchenlin]

@kwea123 Why not colmap init + updating camera parameters during NeRF optimization?

[mwtarnowski]

Done.

[yenchenlin]

Thank you!

[kwea123]

@kwea123 Why not colmap init + updating camera parameters during NeRF optimization?

Yes, it also improves the score, as shown by the paper. It's something I'd like to implement and test in the future.

@yenchenlin btw, I have updated my progress on NeRF-W #24 and #25, please review it when you have time, thanks.

[yenchenlin]

@kwea123 no problem!