Challenges in training on a Custom 3D Face Mesh Dataset

[gokudera3306]

Hi, I am currently attempting to reproduce the results of your paper using a custom dataset. My dataset includes a 3D face mesh with approximately 3000 vertices. However, I have encountered some issues during training and model application that I hope you could help with.

Loss Behavior During Training:

I observed that the loss decreases to its minimum point around 25 epochs without modifying any parameters, but then it starts increasing again. This behavior is perplexing, and I'm unsure whether it indicates overfitting, underfitting, or some other issue.

Model Performance During Editing:

When using the trained model for face editing, the resulting deformations often lead to unrealistic and distorted faces. Specifically, small adjustments can cause significant and unnatural changes to the face geometry, making it difficult to achieve the desired edits.

My training uses the default parameter settings on GitHub.

I would greatly appreciate any insights or recommendations you might have to address these issues. Are there specific adjustments or considerations I should be aware of when training on a dataset with my own data?

Thank you for your time and assistance.

Best regards,

Adam