Question about train.txt and test.txt differences from original HumanML3D

mingyuan-zhang / ReMoDiffuse

ReMoDiffuse: Retrieval-Augmented Motion Diffusion Model

https://mingyuan-zhang.github.io/projects/ReMoDiffuse.html

Other

331 stars 14 forks source link

Question about train.txt and test.txt differences from original HumanML3D #24

Closed kinglr closed 3 weeks ago

kinglr commented 3 weeks ago

Hi, thank you for the amazing work and for sharing this project!

I noticed that your train.txt and test.txt split files differ from those in the original HumanML3D dataset. Could you explain the reason behind these differences?

Thanks again! 😊

kinglr commented 3 weeks ago

If anyone's curious, they simply removed motions shorter than a specified threshold and added the subsequences, rather than processing them on the fly as in here: https://github.com/EricGuo5513/text-to-motion/blob/750f6391bbdbc8bcc6a246d2583f1e6c412d5a25/data/dataset.py#L390.