Inquiry About Potential Release of Training Code

[lauvli]

Dear PartField Team,

First, thank you for sharing this interesting work and making the inference code publicly available. The results are impressive, and I've enjoyed experimenting with the provided demos.

I wanted to kindly inquire whether your team has plans to release the training code for PartField in the future? Understanding the training pipeline would be incredibly valuable for:

Reproducing/extending the research

Adapting the method to new domains

Helping the community build upon this work

If there are no current plans, could you share any insights about potential challenges (e.g., dependencies, data licensing) that might prevent its release?

Thank you for your time and consideration. Looking forward to hearing your thoughts!

Best regards,
lauvli

[Colin97]

The team does not have a clear plan to release the training code in the short term due to limited bandwidth among full-time employees for code cleanup and access limitations for former interns.

However, the full model architecture has been released. If you wish to reproduce the training process, the main challenge lies in preparing the data processing pipeline—preprocessing meshes, rendering them, generating part proposals, etc. These components require careful implementation.

[lauvli]

The team does not have a clear plan to release the training code in the short term due to limited bandwidth among full-time employees for code cleanup and access limitations for former interns.

However, the full model architecture has been released. If you wish to reproduce the training process, the main challenge lies in preparing the data processing pipeline—preprocessing meshes, rendering them, generating part proposals, etc. These components require careful implementation.

I encountered a question while reproducing your work. Has your team considered or compared using PointTransformer, which currently delivers strong performance, as the preceding feature extraction network to replace the PVCNN+Transformer structure? What was the rationale and design considerations behind choosing PVCNN+Transformer as your feature extraction network?

[yxchng]

@Colin97 is there any code that i can refer to regarding the data processing pipeline?

[Colin97]

@lauvli We have not compared against other backbones, but we believe that more recent and stronger backbones capable of encoding a 3D point cloud into a continuous 3D feature field would be beneficial.

[Colin97]

@yxchng Unfortunately, the team has no plans to release the data processing scripts at this time. They do require careful implementation to produce correct results.