Closed chenerg closed 3 years ago
We use PCA based data augmentation strategy to narrow down the rotation space. You can set num_of_transform to 3 or 4 in training, empirically it does not matter a lot and depends on your GPU memory. In the testing stage, the shape should also be aligned with PCA. I will update the testing code for the rotated shapes then. Thanks for your interest.
Hi @NolenChen! How do you perform the shape alignment with PCA? I have set num_of_tranform to 4 in training and I want to test it on the real-world dataset (each object is not well-aligned). How should I perform the alignment? Thank you!
https://github.com/NolenChen/3DStructurePoints/blob/master/evaluate/evaluate_corres_accuracy.py#L84
I have noticed
num_of_transform
, but I don't konw the value for training and test.