facebookresearch / localrf

An algorithm for reconstructing the radiance field of a large-scale scene from a single casually captured video.
MIT License
956 stars 62 forks source link

how to generate the free-view video from your code? #13

Closed Holmes-Alan closed 12 months ago

Holmes-Alan commented 1 year ago

Wonderful work! I wonder how do I use the trajectory path to render a video just like the demo you show in the paper?

edalvb commented 1 year ago

As far as I've come, the videos are generated at the "Optimization" point. You can use the command mentioned by @Yarroudh to reduce times.

Holmes-Alan commented 1 year ago

I understand. However, the command is for training. It is not for video generation. Have you successfully run the command for video rendering?

edalvb commented 1 year ago

Yes, but it also creates a folder called "smooth_spline" and inside it goes the video (video.mp4) that I think you are interested in.

Here is the content of the folder: image

Holmes-Alan commented 1 year ago

Thank you for your answer! I found the training process very slow so I have not seen the video output. I will wait and see. Do you know how to set up the iteration time to reduce the training process?

edalvb commented 1 year ago

I think you could adjust the n_iters_per_frame and n_iters_reg parameters, doing this might reduce the training time. However, keep in mind that it can affect the quality of the results.

ameuleman commented 1 year ago

train.py with --render_path 1 (default) generates the video after the optimisation. If the optimisation ran already, --render_only 1 skips it.