How to use tri-plane to build a radiance field?

[Yiran Xu]

HI there, thanks for releasing the code for this great work!

I have a question about how to use tri-plane representation for general 3D reconstruction task, i.e., building a NeRF. For example, how can I reproduce the result in Figure 3 in the main paper?

I noticed that in Section 4.3 in your supplemental material, you mentioned that you treated the tri-planes as learnable parameters. Is the process like this: the decoder takes tri-planes as the input, and outputs the color and density, and then we do neural rendering to get the 2D image. After we compute some reconstruction loss, we back-propagate it to update tri-planes and the decoder?

I'd really appreciate it if I can hear more from you. Thanks in advance!