Fixed VGGFace2

[yasakova-anastasia]

Summary

• Fixed the problem VGGFace2 as FR dataset #77 (add labels; landmarks/bboxes are optional, but the .csv file with image names must be)

How to test

Checklist

• I submit my changes into the develop branch
• I have added description of my changes into CHANGELOG
• I have updated the documentation accordingly
• I have added tests to cover my changes
• I have linked related issues)

License

• I submit my code changes under the same MIT License that covers the project.
  Feel free to contact the maintainers if that's a concern.
• I have updated the license header for each file (see an example below)

# Copyright (C) 2020 Intel Corporation
#
# SPDX-License-Identifier: MIT

[jihyeonyi]

Most FR researchers usually align and crop faces before the training as a pre-process.
So in the training time, they only use images without the '.csv' file.
So how about generating Lable annotation even if '.csv' file isn't provided.
Or another way to handle this is to create a new extractor for the general FR dataset.

[zhiltsov-max]

@jihyeonyi,

Most FR researchers usually align and crop faces before the training as a pre-process.
So in the training time, they only use images without the '.csv' file.
So how about generating Lable annotation even if '.csv' file isn't provided.
Or another way to handle this is to create a new extractor for the general FR dataset.

Isn't it exactly ImageNet classification format?

[jihyeonyi]

@zhiltsov-max

@jihyeonyi,

Most FR researchers usually align and crop faces before the training as a pre-process.
So in the training time, they only use images without the '.csv' file.
So how about generating Lable annotation even if '.csv' file isn't provided.
Or another way to handle this is to create a new extractor for the general FR dataset.

Isn't it exactly ImageNet classification format?

Technically you're right if ImageNet supports more image formats.
But is it a little bit strange if the user should use ImageNet format for VggFace2?
I'd like to know your opinion.

[zhiltsov-max]

@jihyeonyi, yes, it is strange. Actually, I think, we need to slightly change the structure of Extractor to simplify this kind of integration between sub-formats (ImageNet dataset uses detection from from PASCAL VOC, CamVid uses segmentation from VOC etc.).

[jihyeonyi]

@jihyeonyi, yes, it is strange. Actually, I think, we need to slightly change the structure of Extractor to simplify this kind of integration between sub-formats (ImageNet dataset uses detection from from PASCAL VOC, CamVid uses segmentation from VOC etc.).

Good idea. I'm of the same opinion.
So, we can expect this issue to be solved next time, right?

[zhiltsov-max]

@jihyeonyi, it definitely won't be done in this PR, more likely in Q1.

[zhiltsov-max]

Why there can be no label directory prefix?

[yasakova-anastasia]

If for example item.id = a/1 and there is no label

[yasakova-anastasia]

Maybe add a folder for unlabeled items (like in ImageNet)?

[zhiltsov-max]

I think it is a good idea, because otherwise we require labels.txt.

[zhiltsov-max]

The only problem I see is that the former issue is not resolved.