Datasets compatibility? about 2d-and-3d-face-alignment HOT 7 CLOSED

Hi @hzh8311, The two numbers are used to aproximatevly map the center and the scale obtained from a tight bounding box to the one used during training.

Regarding the results: You should get around 74% if you run this code. For LS3D-W balanced you have to remove the 1.75 shift at this line:

from 2d-and-3d-face-alignment.

So why 1.75 here? Should not it be 1. ?

from 2d-and-3d-face-alignment.

It seems that the center and scale for training are fixed as (225, 275) and 1.8, respectively. And the center and scale for testing are computed according to bounding box. Does this difference matters and leads to such performance drop (17-18%)? @1adrianb

from 2d-and-3d-face-alignment.

@hzh8311 the center is fixed for 300W-LP because the images are already normalised and therefore computing a bounding box is not required. At test time however the faces are coming in all sort of sizes and are found at various locations in the image, therefore some sort of normalisation is required. This is done for simplicity based on the bounding box size.

The error of 17% is not coming from normalisation diff, but from the shift between the ground truth and the predictions. The shift is applied after the predictions are obtained.

from 2d-and-3d-face-alignment.

I removed the shift you mentioned, but only 3-4% improvement compared to my first result. That is to say it is still 13-14% lower than the result reported in the paper.
For more details, the AUCs on traning set and validation set(300W_LP_test) are quiet high (~80%), but it drop badly on LS3D-W dataset.

from 2d-and-3d-face-alignment.

Oh, so you are using a retrained model? Because using the pretained model, already provided, you should be able to match exactly that numbers.

from 2d-and-3d-face-alignment.

Yes, but I can't figure out why the retrained model can not reach your performance.

from 2d-and-3d-face-alignment.