Canonical Volume

[caiyongqi]

Hi. Thanks a lot for providing the code for this work!
I'm a little confused about the canonical volume, is there only one canonical volume modeled for a sequence of images?
NR-NeRF doesn't seem to be able to model large non-rigid motions, such as a scene where a person is dancing.

[edgar-tr]

Hi, yes, that's correct. There is only one canonical volume. Doing so enforces correspondences across the entire sequence, which empirically limits the amount of deformations that can be handled (due to the need for correct long-term correspondences, which is an extremely difficult problem). The upside is that it increases the amount of difference between the input camera trajectory and novel view camera trajectory that can be handled. It's a trade-off between (A) long-term correspondences and interesting novel views and (B) size of deformations. If (B) is more important, concurrent works like Neural Scene Flow Fields (for example) might be better suited.

[caiyongqi]

Thanks!!!