BlazeFace confidence score about pinto_model_zoo

[PINTO0309]

There is a lack of normalization of the input image and a lack of numerical conversion after inference.

img = cv2.imread('2-1.jpg')
img = cv2.resize(img, (128, 128))
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img - 127.5
img = img * 0.007843
image_np_expanded = np.expand_dims(img, axis=0)

(pred_classificators , pred_regressors) = persistent_sess.run([classificators , regressors],
        feed_dict={img_inputs_face: image_np_expanded})

scores = 1.0 / (1.0 + np.exp(-pred_classificators))
print('scores:', scores)
print('pred_regressors:', pred_regressors)

scores: [[[1.92749663e-04]
  [3.74617724e-04]
  [6.53379102e-05]
  [7.41685581e-05]
  [5.91813914e-05]
  [2.26559394e-04]
  [1.27456326e-04]
  [1.55236325e-04]
  [2.33076513e-04]
  [2.11102131e-04]
   : 
   :

pred_regressors: [[[ -1.0449376    0.12775415  21.445286   ...  -1.3779925    2.7564733
    -4.6433153 ]
  [ -0.97221917   0.34412926  26.561253   ...  -0.48758936   3.127716
    -6.060492  ]
  [ -0.5468276   -1.4531596   20.107094   ...  -1.6310003    6.432667
    -6.4479685 ]
  ...
  [ -1.5552132   -6.239525    86.65577    ... -15.01696     12.402348
     3.2532141 ]
  [ -1.292839    -8.317465    96.59374    ... -16.379583    14.803879
     1.740176  ]
  [ -1.857463   -10.383455   107.36553    ... -20.176395    18.376274
    -0.34014758]]]

[sonfire186]

I get very low score of 0.00013611263 .... 8.4253545e-05How to interpret in normal probabilities 0% - 100%?

[Laubeee]

There is a lack of normalization of the input image and a lack of numerical conversion after inference.

img = img - 127.5
img = img * 0.007843

I understand the input has to be normalized [-1 ... 1] not [0...1] (as documented in the model card but to be fair they also are missing the z value in the output..)

May I ask where you got it from? Is this meant to support mean-subtraction, and, in that case, shouldn't the mean values of the training set be used?
In any case, the output I get is quite similar for both ranges..

Edit: The new model you recently added has an additional value for every keypoint... any idea what it is?

[PINTO0309]

I understand that I need to calibrate using the data set used during training. I will follow the exact description in the document if there is one. The normalization range was used in a crude way to predict a likely value. This is because I don't know the correct way to normalize. I don't know the data set either.

If you guys are dissatisfied with the accuracy, you can easily invert the model using the tool below. The dataset used for calibration and the range of normalization can be changed by yourself by adjusting the tool parameters. The last model I converted only used this tool to perform a normalization of RGB/255. But I don't know the certainty of that value.
https://github.com/PINTO0309/tflite2tensorflow

[Laubeee]

I'm not sure I get the concept of what you propose, wouldn't setting custom normalization values have to affect all the model weights too?