About feature score

[mazzzystar]

In the original author's code, they use a FaceModel to rank the features(such as young/smile) between a new sample and the whole datasets.

However, in your code you use simply MSE diff: https://github.com/dsanno/chainer-dfi/blob/888945e9d2297dd535f6ddbe02817b2089d3fdb4/src/train.py#L76

What's your idea on these two method ?

[dsanno]

Thank you for your comment. In "Selecting S^t and S^s" section of original paper, the authors mention two methods to rank images.

• counting thenumber of matching attributes (e.g., matching gender, race,age, hair color)
• cosine distance in deep feature space

I overlooked it and interpolation result may be improved by using one of them.

[mazzzystar]

@dsanno thanks for your quick response. I didn't notice that part in paper.

Actually in author's code they choose the first, they use a pretrained model to score these attributes and rank.

In my own dataset I don't have lots of labels compared with CelebA to train a face attribute model. So I'd like to try the second method. But is that similar with the VGG feature distance ? I worried these feature conficts with the final VGG model and can lead to less meaningful attribute distance. What's your idea on that ?

[dsanno]

@mazzzystar

But is that similar with the VGG feature distance ?

Yes, it is.

I worried these feature conficts with the final VGG model and can lead to less meaningful attribute distance.

I'm not sure about that point. I think your concern is most of nearest neighbor images in deep feature space can be with same attribute as target attribute. But contribution of only one attribute should be small, so I think nearest neighbor images have both same attribute ones and opposite ones as target.