The results measured by the deep3d_pytorch validation set have the mean error 1.87mm and median error 1.78mm

[soubhiksanyal]

The results measured by the deep3d_pytorch validation set have the mean error 1.87mm and median error 1.78mm

[TimoBolkart]

While we have seen validation errors to be 0.1-0.2 mm higher than the actual test set error, a difference of 0.76 mm seems too high. The authors of deep3d_pytorch have provided the test set meshes to us, so I don't know exactly how they obtained these.
Some methods output only a tight face crop of BFM meshes, which results in a higher error in the forehead and close to the face boundary. Instead, providing the uncropped BFM meshes reduces the error. But I don't know if this is what deep3d_pytorch people did.

[DougDong666]

The deep3d_pytorch use the cropped mesh, which does not contain the jaw. And this is a big region which makes the mean error and median error so big. when other people test the deep3d_pytorch, I suggest you can only get First 80 dimension output parameter, which is the id parameter, to generate the non-cropped mesh.

[DougDong666]

the mean error of non-cropped mesh is 1.39mm and the median error of it is 1.27mm

[Oliver-YX]

The deep3d_pytorch use the cropped mesh, which does not contain the jaw. And this is a big region which makes the mean error and median error so big. when other people test the deep3d_pytorch, I suggest you can only get First 80 dimension output parameter, which is the id parameter, to generate the non-cropped mesh.

I wonder how to only get the first 80 dimension output parameters when we test the deep3d_pytorch. I meet the same problem and the result is 1.38mm mean error and 2.02mm median error. Thanks for your answer