[Feature] Support single depth to 3D hand keypoints baseline, NYU Hand Pose Dataset and AWR

[walsvid]

Motivation

Integrate single depth to 3D hand keypoints into mmpose.

Modification

Add an algorithm for estimating 3D gestures from a single depth image to mmpose
Includes the following changes:

• Add 3d_kpt_sview_depth_img common dataset NYU Hand Pose dataset
• Add 3d_kpt_sview_depth_img basic baseline algorithm AWR (AAAI2020) network architecture
• Add data transformation pipeline for depth image data pre-processing

🚧 The evaluation code is still WIP, code review are required for above part.
🚨 More unit test are needed. Maybe it could be discussed during the code review stage.

BC-breaking (Optional)

• Modified TopDownGenerateTargetRegression to adapt the calculation method of the GT target coordinates of the algorithm that uses the depth image as input

Use cases (Optional)

Checklist

Before PR:

• I have read and followed the workflow indicated in the CONTRIBUTING.md to create this PR.
• Pre-commit or linting tools indicated in CONTRIBUTING.md are used to fix the potential lint issues.
• Bug fixes are covered by unit tests, the case that causes the bug should be added in the unit tests.
• New functionalities are covered by complete unit tests. If not, please add more unit tests to ensure correctness.
• The documentation has been modified accordingly, including docstring or example tutorials.

After PR:

• CLA has been signed and all committers have signed the CLA in this PR.

[CLAassistant]

All committers have signed the CLA.

[jin-s13]

Thanks a lot for your contribution! We will review the PR ASAP.
@walsvid Please sign the CLA so that we can proceed the code merging.

[walsvid]

Hi, @jin-s13 . I have signed the CLA. I will continuously submit some commits in this branch recently to improve the PR.

[walsvid]

Code will be significantly refactored to fit v1.x, so it will take some extra time.

[ly015]

Code will be significantly refactored to fit v1.x, so it will take some extra time.

Thank you very much. We will maintain both master and 1.x branches until 2023, and still release v0.x versions with new features in this period. So it would be great to create a new pull request for the dev-1.x branch to add this feature to 1.x version. And this PR will still be merged into the master branch after review.

[jin-s13]

Hi @walsvid, hope you're well. Thank you again for the contribution!

Is this PR ready for review? Or shall we wait for your further updates/commits before code review?

BTW, have you already trained a few models on NYU dataset? Would you like to share the checkpoints and training logs with us?

[walsvid]

Hi, I think this PR still need updates for code review. As for NYU dataset, I have checkpoints of AWR model, I can provide link. But the log is messy, this may not help. I'll update thie PR ASAP.

[jin-s13]

Hi, I think this PR still need updates for code review. As for NYU dataset, I have checkpoints of AWR model, I can provide link. But the log is messy, this may not help. I'll update thie PR ASAP.

Well understood. Please let us know if you have met any problems/questions. Thanks.

[Enderfga]

Hi @walsvid ,I'm doing some nyu hand dataset and awr related research and would like to use mmpose, would you be so kind as to provide the code and checkpoints, thank you very much!

[walsvid]

Hi @walsvid ,I'm doing some nyu hand dataset and awr related research and would like to use mmpose, would you be so kind as to provide the code and checkpoints, thank you very much!

Hi, @Enderfga @jin-s13. I upload the ckpt and nyu data meets the coco format annotations to Google Drive. The current Pull Request still needs to improve the code related to evaluation. It would be better if @Enderfga can help to improve.