Totoro97 / f2-nerf

Fast neural radiance field training with free camera trajectories
https://totoro97.github.io/projects/f2-nerf/
Apache License 2.0
933 stars 69 forks source link

Problems training model from UnrealEngine data generation #66

Open mgrova opened 1 year ago

mgrova commented 1 year ago

Hello, first of all thanks for sharing you work!

I am trying to train the model with an UnrealEngine environment and getting the images and pose information through AirSim. That way we simulate that the images are taken by a drone that is flying over an area. The idea with this is to try to generate images from unseen points of view.

The camera pose information, I am getting it in FLU(x:front, y:left, z:up) in reference to an origin set in the unreal environment. Through the code snippet you passed in this issue, #29, I am converting this pose information to the "cams_meta.npy file" (Converting the pose from FLU to OpenGL style (I hope I'm doing that conversion right hahaha)). And getting the bounds of the simulated depth camera.

When I use the generated data, during training I get the ouput 'nan!' constantly. But, if I use the pose data obtained through colmap it works correctly (and using the following script: colmap2poses). In the following shared folder there are the files: shared folder.

Im using the 'wanjinyou' config file by default. Should I change the bounds_factor of the dataset? Can this problem be due to the scale?

Thanks you in advance.