Hierarchical Labeling

[bonhunko]

Summary

Resolves 93482

• Add LabelGroup option to datumaro for representing 'exclusive'

How to test

1. Generate a dataset containing hierarchical labels.
2. Export the dataset in 'datumaro' format.
3. The labels which disobey the single-selection rule will be filtered-out in the exported 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

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[vinnamkim]

LGTM.
But I don't still get the concept of this work. Our current data structure of categories has a parent entity. It means that it can represent any tree structure with arbitrary depth and width.

How about this case?

        l1(T)
      /        \
 l2(F)        l3(F)
  /  \         /    \
l4    l5     l6     l7

and Item1.annotations := [l1, l2, l3, l4, l5, l6, l7] (if label has single_selection=True, then it denotes label(T), otherwise, label(F)). Then, this filtering will make Item1.annotations = [l1, l2, l4, l5, l6, l7].

I feel this is a bit inconsistent because l6 and l7 is still alive although their parent l3 is filtered out.

[wonjuleee]

LGTM

[wonjuleee]

Could you check whether Vinnam's concern is true or not?

[bonhunko]

Good, questions. This is a kind of architecture-level question. But I think that datumaro is needed to exports a general output. The detailed artchitecture-level filtering should be done by customer side. Because the detailed policy about how to handle hierarchical labeling should be determined by the customer. So please think that this implement only filter-out the error-cases that obviously wrong.

After facing some practical cases, we may change this filtering policy again.

[bonhunko]

As vinnam suggested and discussed with wonju, I'll follow the LabelGroup implementation to mark the 'exclusive' after that I'll request review again.

[bonhunko]

As we discussed, I import the LabelGroup format from the OTX, please review again.

[vinnamkim]

LGTM.