Open Images: implement bounding box support

[IRDonch]

Summary

This continues the implementation of #274.

How to test

You can download some of the bounding boxes for OID, create images.meta as described in the documentation and try to import the dataset.

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) 2021 Intel Corporation
#
# SPDX-License-Identifier: MIT

[zhiltsov-max]

Also add a test like in #346

[IRDonch]

Also add a test like in #346

I have examined the implementation in #346, and it would be tricky to implement something like that for OID, because the annotation files often don't have fixed names - just possible suffixes. So it wouldn't be possible to specifically remove the annotation files for a particular subset that was removed. Would it work if I just removed all existing annotation files at the beginning of apply?

[zhiltsov-max]

Honestly, it is tricky every time. We know from the patch which items were added, removed or modified and which subsets were modified. During writing it is also easy to find out modified subsets that became empty now. Will this information be enough to remove unnecessary files in this format?

[IRDonch]

We know from the patch which items were added, removed or modified and which subsets were modified. During writing it is also easy to find out modified subsets that became empty now. Will this information be enough to remove unnecessary files in this format?

The problem is that we don't know which files which annotations/metadata came from. For the training subset, it could be train-images-with-labels-with-rotation.csv, oidv6-train-images-with-labels-with-rotation.csv, image_ids_and_rotation.csv, or (if the user tries to really exercise the flexibility of the loader) even foobar-images-with-rotation.csv. So I don't see a better solution than just nuking the annotation directory.

Although I guess for safety I could do the nuking after the new files are written, instead of before.

[IRDonch]

Anyway, I think patching support is a distinct issue from bounding box support, so I think I'm going to work on it separately.

[zhiltsov-max]

I think, the severity argument should be separate from the error class. Probably, I see reporting a dataset problem as an Extractor (Converter) - base class - method, which could have such severity parameter. Then, it could be handled according to the current error control policy - e.g. a warning could be promoted to an error or ignored. #142. It is a big new feature, though.

[IRDonch]

I don't understand - do you want me to change something in the code? I didn't implement the UndefinedLabel class, I'm just using it.

[zhiltsov-max]

Probably, this error class is not supposed to be used outside dataset validators. But we can leave it.