Unbroken models about multi_part_assembly

[Wuziyi616]

@YunChunChen pining my teammate on this. IIRC we use the object name as the directory name of broken parts. So there should be a way to find the original 3D model. But we cannot share the original data due to license issue. So probably you have to download those datasets (PartNet and Thingi10K) and retrieve from them.

[YunChunChen]

You can get the original mesh file name by parsing the folder name. For example, if the folder name is everyday/Bottle/7984d4980d5b07bceba393d429f71de3/, then the original mesh file should be something like Bottle/7984d4980d5b07bceba393d429f71de3.obj

[Wuziyi616]

I think we're actually using a variant of PartNet right? From a robotics grasping dataset called ACRONYM. This is because the shapes in this dataset are aligned to a canonical pose.

For Thingi10K it's a bit complicated, you should go to their webpage and search for the object id. For example, you can get the original of artifact/39087_sf/ by searching for id=39087_sf on the website. I believe they have API or script for batch automatic downloading.

[Wuziyi616]

Feel free to re-open the issue if you have any further questions :)

[Pterosaur-Yao]

Thank you for the valuable information. However I still have some doubts, if I want to get the original mesh files of the EVERYDAY dataset, should I just download the PartNet dataset? I find that PartNet offers three versions(datav0, sem_seg_h5, and ins_seg_h5), which one should I download?

[Wuziyi616]

We use a PartNet variant called ACRONYM. You probably have to download it. I currently don't have access to this original dataset on the cluster, so I cannot confirm if the ID is exactly the same or under some naming conversion. But I think they should be the same.