martinnormark
commented
1 month ago I am not associated with the team who wrote the paper and is not at all expert in 3D, meshes etc. - I found the paper, read it many times and decided to try to implement it.
I have not completed the implementation, I have written a yolo training loop but not tried to run it yet 🙈
From what I understand and my own intuition, the model may converge on out-of-distribution data from its training strategy of using Point Sampler, Edge Predictor and Face Classifier in a way that:
- Uses local geometric features
- Avoids global topology assumptions
- Learns generic simplification operations
- Validates on different mesh types
It would probably struggle on meshes that are very complex or highly irregular compared to the training data. For example, the ABC dataset I am using here is relatively simple compared to e.g. game character meshes. So the trained weights from the industrial CAD meshes would probably not do well in gaming, and thus needs to be re-trained on the domain it needs to be applied in.
On the training part, the meshes, epochs and batch size are not connected in that way. Training happens in a number of loops (epochs) which will traverse items in the training set using a batch size. Read more here: https://machinelearningmastery.com/difference-between-a-batch-and-an-epoch/
SMY19999
Hi!
Thanks for your great contribution. It is really a novel method that is differential and fast!
I have read the paper and the code.
My question is how to train the model?
The number of vertex of the mesh in the dataset is not the same.
Given a mesh, has N vertices, so is the batch size N?
Do you train one mesh in one epoch? This means the epoch loss is only about this mesh.
If I'm right, I wonder that how the model can generate to out-of-training data with the different number of vertex and different topology?
Look forward to your reply. I'm really want to learn from this work!