about the PCA on the training set

[justusschock]

about the PCA on the training set

[justusschock]

I don't really understand your problem. The basic pipeline is:

1. Create a PCA on the trainset's landmarks
2. Use this PCA (only the first N compontens + mean) inside the shape model
3. Learn the weighting parameters and global transformation parameters by CNN

An in prediction:

1. Use trained CNN to predict parameters and apply PCA-Layer and global transformation to obtain shapes

[zrqi]

I mainly want to ask what is the label when training the cnn? Is it the weight coefficient ？How to calculate it based on the training sample

[justusschock]

No, the label are the coordinates of the groundtruth points. We use a simple distance loss (L1) between the predicted point coordinates and the groundtruth coordinates, which is implicitly mapped on the weight coefficients by the proposed layer.

[zrqi]

Thank you very much, I understand.