search

markomih / ResFields

[ICLR 2024 Spotlight ✨] ResFields: Residual Neural Fields for Spatiotemporal Signals
https://markomih.github.io/ResFields
MIT License
162 stars 11 forks source link

ResFields + Dynamic Point Field? #8

Closed fishfishson closed 3 months ago

fishfishson commented 4 months ago

Dear author,

Thx for sharing your great repo! I wonder whether ResFields can be applied to Dynamic Point Field? Maybe you have already tried it. Can you share your insights about it?

markomih commented 4 months ago

Hi @fishfishson,

Thanks for your interest in our work!

Indeed we have tried DPF + ResFields for learning scene flows. Please see Section 4.4 in our paper for more details. On our setup (Tab. 5), we demonstrated that ResFields benefit learning scene flows, i.e. a smaller architecture with ResFields (128 neurons) achieves better results compared to a larger DPF (256 neurons).

Best, Marko

fishfishson commented 4 months ago

Hi, Marko

How about using DPF + ResField for dynmaic scene reconstruction or rendering?

markomih commented 3 months ago

Hi @fishfishson,

We haven't directly tried DPF + resfields for the dynmaic scene reconstruction as it's unclear how to utilize DPF for 4D NeRFs.

Best, Marko