Bad 3DGS reconstruction

[andrewsilver1997]

Hi,

Very interesting work. I'm running your work on my self-customized data. I observed that the result of python train_gaussian.py -s <path to COLMAP or NeRF Synthetic dataset> wasn't very good compared to the vanilla 3DGS. And the renderings look like a group of spheres. Is it normal and are there any changes in your code?

And for convert_data.py, how many folders do I expect after running it. I noticed for training stage 1, it requires multi gaussian.npy and occ.npy. How is that possible for only one volume? Or the idea is to train a vanilla 3DGS model for different volume and then feed all trained 3DGSs to VAE?

Thank you