API & Code Cleanup

[martinheidegger]

Result of working on first dockerized version of KYCDeepFace and other refactorings.

Contains:

• 1-to-1 API

  https://github.com/tradle/KYCDeepFace/blob/69b00018fd5446f755bdff3d9fd37a6f103a037d/vision/kycdeepface.py#L53 https://github.com/tradle/KYCDeepFace/blob/69b00018fd5446f755bdff3d9fd37a6f103a037d/vision/kycdeepface.py#L112

• lambda docker API

  https://github.com/tradle/KYCDeepFace/blob/69b00018fd5446f755bdff3d9fd37a6f103a037d/%CE%BB.py#L33-L38

• AWS Deployment

  .github/workflows/deploy.yml

  which actively deploys to AWS

• Refactoring of configuration of Image Processing configuration

  https://github.com/tradle/KYCDeepFace/blob/69b00018fd5446f755bdff3d9fd37a6f103a037d/config.py#L58-L69

• Refactored first two CLIs:

  • bin/pytorch-to-onnx.py
  • bin/webcam-alignment-example.py

• First set of unit tests test.py (executed as github actions)

• Contains the first set of models for tests.

[martinheidegger]

I extracted this new face alignment variant from @fyr91's face alignment code that should be added to this PR:

import skimage
import numpy
import cv2

# reference - by default 112x112 input size
glint_align_dst = numpy.array(
    [
         [30.2946, 51.6963],
         [65.5318, 51.5014],
         [48.0252, 71.7366],
         [33.5493, 92.3655],
         [62.7299, 92.2041]
    ],
    dtype=numpy.float32
)

# move all points 8 pixel to the right
glint_align_dst[:,0] += 8.0

def glint_align(image, landmarks):
    # makesure imput image size is 112x112
    # resize image to 112x112
    # image = cv2.resize(image, (112,112))
    # other wise if 112x96, need to comment out 8 pixel shift in the reference pts
    rightEyePts = landmarks[36:42]
    leftEyePts = landmarks[42:48]
    nose_tip = landmarks[33]
    mouth_left = landmarks[48]
    mouth_right = landmarks[52]
    leftEyeCenter = leftEyePts.mean(axis=0).astype("int")
    rightEyeCenter = rightEyePts.mean(axis=0).astype("int")

    src = numpy.asarray([rightEyeCenter, leftEyeCenter, nose_tip, mouth_left, mouth_right], dtype=numpy.float32)

    # https://scikit-image.org/docs/stable/api/skimage.transform.html#skimage.transform.SimilarityTransform
    tform = skimage.transform.SimilarityTransform()
    tform.estimate(src, glint_align_dst)
    M = tform.params[0:2, :]

    image = cv2.warpAffine(image, M, (112,112), borderValue=0.0)

    return image

@fyr91 I am looking at the destination points and I am wondering why the eyes/mouth points are not level (same y coordinate), wouldn't an algorithm like this be reasonable?

import numpy

def glint_align_dst (width, height):
    center_x = width / 2

    eye_left = center_x - height * 0.1585
    eye_right = width - eye_left
    eye_y = height * 0.46

    nose_x = center_x
    nose_y = height * 0.64

    mouth_y = height * 0.8241
    mouth_left =  center_x - height * 0.13
    mouth_right = width - mouth_left

    return numpy.array(
        [
            [eye_left, eye_y],
            [eye_right, eye_y],
            [nose_x, nose_y],
            [mouth_left, mouth_y],
            [mouth_right, mouth_y]
        ],
        dtype=numpy.float32
    )

@fyr91 this alignment is seems very different from the previous alignments. Before the left eye was at 28.8 x 33.6 now it is at 30.2946 x 51.6963; Seems like with this we need to (re) train the model again, right?