Missing PartNet-Mobility Train/Test Split Files & Clarification on SINGAPO Filtering

[XIAOTIANCRAZY]

Hi there,

Thank you for your great work and for releasing the code. I am currently trying to reproduce the experiments in the paper.

However, I encountered an issue regarding the dataset preparation. I noticed that the repository does not seem to include the train/test split files for the PartNet-Mobility dataset.

In the paper, it mentions: "For [59], we follow the train-test split by [30]."

While trying to map this to the SINGAPO (https://github.com/3dlg-hcvc/singapo/blob/main/data/data_split.json) setup, I observed that:

1. The original filtered PartNet-Mobility contains 566 objects.
2. The main experiments cover 7 categories:
   • Storage
   • Table
   • Refrigerator
   • Dishwasher
   • Oven
   • Washer
   • Microwave

Since the full PartNet-Mobility dataset has many more categories, I am a bit confused about the exact filtering and splitting process.

Could you please clarify:

1. How exactly the 566 objects and 7 categories were selected?
2. How the train/validation/test splits were performed based on [30]?

Most importantly, would it be possible to release the annotation files (e.g., .json or .txt files) that contain the specific train/val split used in your experiments?

This would be extremely helpful for reproducibility. Thank you very much for your time and assistance!

[RuiningLi]

Hi!

Sorry this hasn't been super clear in the paper: Particulate is trained on all categories from PartNet-Mobility except for the ones included in the test split in [30]. So training uses more objects than the training set in [30]. Please refer to the split file below.

data_split_sub16.json