Missing the nerf synthetic dataloader in the sugar train script.

[shawLyu]

Thanks for the great work. I found that when I tried to train the sugar model for the NeRF synthetic dataset, the path was wrong. It seems that the original codebase only supports the COLMAP dataset. Do you have the dataloader for the nerf synthetic dataset? Looking for your reply, thanks.

[Anttwo]

Hello shawLyu,

Indeed, as we focused on reconstructing real scenes, the code lacks a loader for synthetic dataset. I will fix that soon. Could you paste your error in here?

I already know about a problem with scenes from the synthetic dataset, which is that there is no SfM point cloud for them. Authors from the original Gaussian Splatting paper propose to use a random point cloud instead, so I am definitely going to add this option to the code.

[shawLyu]

Thank you for your prompt response.

The error originated from an incorrect dataset structure. In the case of the NeRF synthetic dataset, images are stored in the train/val/test directories rather than the images directory.

Following the instructions in the README, we can initiate training with the vanilla Gaussian and use these points as the starting position for optimizing SuGaR. In my opinion, this approach may have a reduced impact on the final results.

Could you provide any tips or tricks for loading synthetic data? If you don't currently have the corresponding code available, I am willing to attempt implementing it independently.

I appreciate your response and commendable efforts once again!