Add ARM64 support

[NHellFire]

This upgrades libraries to versions that have arm64 wheels available, and builds the remaining ones (imagecodecs) from source. I've only build-tested this as I currently don't have a spare arm64 device to test on, but the test suite passed:

#22 120.0 ==== 39 passed, 1 skipped, 45 deselected, 18 warnings in 101.35s (0:01:41) =====

If anyone wants to test this, you can build the images natively with make from the embedding-calculator folder. Or cross-build and export the image using:

docker buildx build -t compreface-core:arm64 --platform linux/arm64 --progress plain .
docker image save -o compreface-core-arm64.tar compreface-core:arm64

And import with:

docker image load -i compreface-core-arm64.tar

Fixes #610, possibly also #519 (I haven't tested building the CUDA images yet).

Marking as draft for now, as I still need to update the Makefile to use buildkit and try building the CUDA images. I'll also look at building imagecodecs in a separate container to reduce the final image size.

[github-actions]

CLA Assistant Lite bot All contributors have signed the CLA ✍️

[NHellFire]

I have read the CLA Document and I hereby sign the CLA

[pospielov]

Hi @NHellFire, thanks for this fantastic contribution!
The main concern is that the new version should have backward compatibility with the new one.
I mean that the embeddings that compreface-core returns should be the same in the current CompreFace version and after your changes. In this way, old installations won't be broken.
The easiest way to check it is to install the current CompreFace version, upload one image to face collection, and test on another image (preferably on another person), and write down the similarity. Then, install your version and use the same volume for DB. And test the same image on the same service. The similarity should be the same.
Could you do this?

[NHellFire]

Using a v1.1.0 database with my branch produced an identical recognition result for these test images.

Images used:

• Face collection: https://pxhere.com/en/photo/601668
• Test:
  • https://pxhere.com/en/photo/980074
  • https://pxhere.com/en/photo/764654
  • https://pxhere.com/en/photo/418995

Results for the first test image are:

v1.1.0 (x86_64 host):

Similarity: 0.17
Male, Age: 25-32

{
  "result": [ {
      "age": {
        "probability": 0.9716785550117493,
        "high": 32,
        "low": 25
      },
      "gender": {
        "probability": 1,
        "value": "male"
      },
      "box": {
        "probability": 0.99986,
        "x_max": 1863,
        "y_max": 1202,
        "x_min": 1012,
        "y_min": 218
      },
      "subjects": [ {
          "subject": "A",
          "similarity": 0.16651
        } ],
      "landmarks": [ [ 1324,
          643 ],
        [ 1591,
          652 ],
        [ 1478,
          814 ],
        [ 1328,
          911 ],
        [ 1575,
          915 ] ]
    } ]
}

feature/arm64 (x86_64 host, same database):

Similarity: 0.17
Male, Age: 25-32

{
  "result": [ {
      "age": {
        "probability": 0.9716785550117493,
        "high": 32,
        "low": 25
      },
      "gender": {
        "probability": 1,
        "value": "male"
      },
      "box": {
        "probability": 0.99986,
        "x_max": 1863,
        "y_max": 1202,
        "x_min": 1012,
        "y_min": 218
      },
      "subjects": [ {
          "subject": "A",
          "similarity": 0.16651
        } ],
      "landmarks": [ [ 1324,
          643 ],
        [ 1591,
          652 ],
        [ 1478,
          814 ],
        [ 1328,
          911 ],
        [ 1575,
          915 ] ]
    } ]
}