SimilarityTransform returning identity about face-alignment-at-3000fps HOT 4 OPEN

1. scale and rotate can differ very much, and won't always be identity or close to 1 unless your shape is similar to mean shape.
2. you can calculate the mean_shape before any data augment which will avoid the vertical problem you mentioned.

from face-alignment-at-3000fps.

Actually, I wonder why we need such a function i.e. calcSimilarityTransform(), I don't know why we should get a relative delta_shapes for predicting target. Sorry for my silly question.

I can't understand why we need to get a relative delta_shapes for predicting target. I think we can use absolute delta_shape directly for target.
// Get relative delta_shapes for predicting target
vector getDeltaShapes(vector &gt_shapes, vector &current_shapes

from face-alignment-at-3000fps.

@mariolew SimilarityTransform is needed for some reason that other papers are already said.

And for relative delta shapes, it should be relative not absolute. Imagine that you have a face for different size (100x100, 400x400, 600x600 for example). Every face has a shape on it, with extracted features, if delta shape is relative, this three faces will be given the same delta shape which is what we want, but if delta shape is absolute, delta shape can not fit to three different scale unless the feature from these faces are very different which is not what we want to see.

Of course, if you resize all face (training data and test data) to the same size, relative or absolute delta shape doesn't matter I think :)

from face-alignment-at-3000fps.

About the necessity of similarity transform, can you please provide the name of the related papers?

from face-alignment-at-3000fps.