train nerf with custom data

[w2kun]

Hi,

I noticed that the datasets supported in this nerf implementation have been processed in advance. I want to use my own data to train nerf. Can you provide some help on how to prepare custom training data?

[patricklabatut]

Better support for NeRF-style datasets is on our radar. @bottler could likely elaborate on how to prepare your own training data for now.

[bottler]

Pick a name for your scene, e.g. myscene.

If you have N images of shape [W,H], they should be concatenated and saved into a single very tall png of shape [W,H*N] with name data/myscene.png.

The cameras and data splits should be put into a dict as follows and torch.saved into a file data/myscene.pth.

• "cameras". This should contain camera intrinsics in PyTorch3D format, as a dictionary of the following things
  • "R": camera rotation as a FloatTensor of shape [N,3,3]
  • "T": camera position as a FloatTensor of shape [N,3]
  • "focal_length": a FloatTensor of shape [N,2]
  • "principal_point": a FloatTensor of shape [N,2]
• "splits". This should be a list with three LongTensors (or lists of ints) containing numbers between 0 and N - 1 which represent the indices of
  • members of the training set
  • members of the validation set
  • members of the testing set set

[w2kun]

Thanks for the reply.