search

facebookresearch / localrf

An algorithm for reconstructing the radiance field of a large-scale scene from a single casually captured video.
MIT License
973 stars 61 forks source link

How to render the results of training in a 3D scene #15

Open Dragonkingpan opened 1 year ago

Dragonkingpan commented 1 year ago

I ran forest1 on the A6000 for three days and got a video that looks good. At least dragging forest1/transform.json directly into instant ngp is very poor. What I want to know is whether this result can be rendered into a true 3D, the kind that can control the camera movement with the mouse and watch it freely. Thank you!

ameuleman commented 1 year ago

Hi, With the current backbone, rendering is too slow (about 10~20s/frame). An alternative could be to read and render from a pose file, which I can implement if you are interested.

xiemeilong commented 1 year ago

@ameuleman Do you mean the trained TensoRF partition can only be rendered at a speed of 10-20S/frame?

ameuleman commented 1 year ago

With our resolutions, number of samples, and blending between renders (many frames need two renders), yes.