Couldn't reproduce the results of RobustNeRF in the paper

[otonari726]

@Sarasra

Thank you for your support and contribution.

I have been conducting experiments with RobustNeRF loss on the Android and Statue scenes, including images with transient objects. Regrettably, I've encountered challenges in reproducing the results presented in the paper using the config file provided, as shown below. Could you provide guidance or instructions on replicating the results from the paper?

Additionally, I have observed that the camera parameters for the Crab scene appear to be absent. Could you also provide those? This information would be invaluable for my efforts to replicate the experiment.

Thank you again!

my result:
Android (left: original image, right: rendered image)

Statue (left: original image, right: rendered image)

[mirlansmind]

Try changing GLO embedding parameter from 0 to 4 and run the experiment all the way to 250,000 steps. Although there are still artifacts, it is much better than what you get.

[otonari726]

Thank you! I'll try it.
Could you tell me the value of num_glo_embeddings? default(=1000)?