WIP - feat: 🚀 initial keypoint support for transformers added

[onuralpszr]

Description - Working in Progress

Recent release of transformers introduce keypoint support and initial support for keypoint added.

Type of change

• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

Docs

• Docs updated? What were the changes:

[onuralpszr]

@merveenoyan 👋 hello I notice current latest and dev doesn't have "post_process_keypoint_detection" function and I was able to run via only this pull request branch "huggingface/transformers#33200" Today I read about "https://huggingface.co/docs/transformers/tasks/keypoint_detection" you posted in social and I was unable run keypoint but I was able to do via using install this repo/branch, I think it is forgot to merge ?

!pip install git+https://github.com/sbucaille/transformers@superpoint_fix -q

Error output when I ran from "main"

    outputs = processor.post_process_keypoint_detection(outputs, image_sizes)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'SuperPointImageProcessor' object has no attribute 'post_process_keypoint_detection'

Here is my initial collab link for working version

https://colab.research.google.com/drive/1gxpE9iDR7gXl2VwwScB5W8Yw-e_xUo3i?usp=sharing

[onuralpszr]

@LinasKo missing function basically converting normalize values to w,h values to position correctly should I finish based on that in current release and finish the PR ?

[LinasKo]

@onuralpszr, I need more details.

1. Missing function where? Our repo or theirs?
2. "should I finish based on that in current release and finish the PR ?" - Based on what? Based on the function being missing? I'd like more clarity here 😉

[onuralpszr]

@onuralpszr, I need more details.

1. Missing function where? Our repo or theirs?
2. "should I finish based on that in current release and finish the PR ?" - Based on what? Based on the function being missing? I'd like more clarity here 😉

Sorry, I was little bit hasty I guess and for answering your questions

1 - Missing function is theirs and it is post process function for converting normalize xy values to pixel based on xy values (post_process_keypoint_detection is missing)
2 - Currently latest version of transfomers has SuperPointImageProcessor but, doesn't have SuperPointImageProcessor.post_process_keypoint_detection() so I will finish PR based on latest version of transformers to and I will write the helper function to convert image size correction or I had to assume it does use function and leave as it is. You can check the PR link I posted above to see it.