Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Custom dataset #8

Open
damonftl opened this issue Jun 27, 2022 · 4 comments
Open

Custom dataset #8

damonftl opened this issue Jun 27, 2022 · 4 comments

Comments

@damonftl
Copy link

Hi, I'd like to try and make my own dataset of different buildings with parts other than the ones included with BuildingNet. Generating the 3D models (and point clouds) themselves shouldn't be a problem - it's everything else I'm not sure about.

For the minkowski pretrained features, should I simply download their code (https://github.com/NVIDIA/MinkowskiEngine) and run the models through it?

The required json files for each model used to train BuildingNet are fairly extensive (adjacency, containment, support, and similarity) and they don't seem especially trivial to recreate. The code for their creation doesn't appear to be in this repository, and I'm guessing it is part of the labeling application shown in the paper. Can the code for the labeling application be downloaded? I'm hoping in the end to output the label data automatically together with the 3D models when they are generated (that is, a synthetic training data generation pipeline) and being able to see some of that code would be a huge help.

@mzillag
Copy link

mzillag commented Jan 30, 2023

Hi, I'd like to try and make my own dataset of different buildings with parts other than the ones included with BuildingNet. Generating the 3D models (and point clouds) themselves shouldn't be a problem - it's everything else I'm not sure about.

For the minkowski pretrained features, should I simply download their code (https://github.com/NVIDIA/MinkowskiEngine) and run the models through it?

The required json files for each model used to train BuildingNet are fairly extensive (adjacency, containment, support, and similarity) and they don't seem especially trivial to recreate. The code for their creation doesn't appear to be in this repository, and I'm guessing it is part of the labeling application shown in the paper. Can the code for the labeling application be downloaded? I'm hoping in the end to output the label data automatically together with the 3D models when they are generated (that is, a synthetic training data generation pipeline) and being able to see some of that code would be a huge help.

Hi, did you make your pipline? I have the task whick look similar to yours. I need just to put the 3D mesh or point cloud to network and get labeling, that's all, but currently seacrhing for the way how to do it with existing model and code.

@damonftl
Copy link
Author

damonftl commented Feb 2, 2023

Hello mzillag,
I did not end up making the pipeline back at the time I originally posted this. However, I did contact the authors of the paper and they have since uploaded more of the code required here to GitHub. The feature scripts and annotation tool are here now, which should make it possible to use your own custom data. Unfortunately the project I was going to use this for ran out of time before I could really try it.

However, I have a new project starting very soon that will use this, so I'll be getting back to it and making another attempt at using custom data. If I run into anything I'll be sure to post it here.

@mzillag
Copy link

mzillag commented Feb 2, 2023

Thanks for the answer. So you didn’t know how to test the model on point cloud (which has not labeled) not from their dataset?

@damonftl
Copy link
Author

damonftl commented Feb 2, 2023

That's right, exactly. I was able to train and test fine with their dataset, but not do either with my own. Hopefully soon though!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants