Questions about reproducing Replica example and PyTorch log_prob_unnorm method

[Tripton]

I am currently working on integrating my own RGBD dataset, which I previously used for normal Gaussian Splatting without depth, into this repository. For this, I'm investigating how to handle depth, including loading and scaling depth, creating the raw point cloud (raw_pc), and so on.

As I still struggle with it, I am trying to reproduce the project using your provided Replica Dataset example. However, I have encountered a couple of issues:

1. Replica Dataset Example: I cannot locate the traj.txt file which seems to be necessary for the example. Could you please upload it or direct me to where I can find it?

2. PyTorch issue with MultivariateNormal: While experimenting, I ran into an issue with this line of code . It seems that PyTorch does not have a log_prob_unnorm method for MultivariateNormal. I am uncertain if it is correct to use log_prob instead. Do you have any suggestions on how to fix this or what the appropriate method would be?

Thank you in advance for your help!

[siyanhu]

I solved it by using colmap to generate sparse data...

No you don't have. Dataset should be here: https://github.com/facebookresearch/Replica-Dataset

[kungfrank]

Sorry for the late reply. I've updated the google drive folder!

[kungfrank]

I believe the second issue is caused by a different pytorch version. I'm using 1.12.1. You can do

import torch
print(torch.__version__)

to check which version you are using.

[mappro6]

@kungfrank Hi, I use the pytorch 1.12.1, but the MultivariateNormal still does not have a log_prob_unnorm method and only log_prob can be used. Is log_prob correct for training?

[kungfrank]

Hi @mappro6, Thank you for reporting this. The log_prob_unnorm is a custom function. Please pull the repo again to get the updated code.