alexvbogdan/DeepCalib

Question about preprocess

Closed this issue · 1 comments

iwatake2222 commented

Thank you for sharing the great project and the models.

I have a question about the preprocess.

DeepCalib/prediction/Regression/Single_net/predict_regressor_dist_focal_to_textfile.py

Line 94 in b9b5fc3

image = preprocess_input(image)

It looks two preprocessing are executed:

  • normalize the input data from [0, 255] to [-1.0, 1.0]
    • L89 - L91
  • then, keras.applications.imagenet_utils.preprocess_input is called
    • L94

As a result, the input tensor data becomes like the following:

array([[[[-104.92331373, -117.62998039, -124.35843137],
         [-104.939     , -117.67703922, -124.4054902 ],
         [-104.89194118, -117.59076471, -124.28      ],
         ...,
         [-104.93115686, -117.74762745, -124.60941176],
         [-104.93115686, -117.76331373, -124.6172549 ],
         [-104.939     , -117.75547059, -124.6172549 ]],

        [[-104.92331373, -117.63782353, -124.35843137],
         [-104.92331373, -117.64566667, -124.37411765],
         [-104.90762745, -117.59860784, -124.28784314],
         ...,
         [-104.939     , -117.75547059, -124.6172549 ],

Is this what you intended?
I just use the same preprocess as above and it works somehow.

By the way, I made a project to run the model on Android with TesorFlow Lite.

Thanks!

alexvbogdan commented

Yes, that is how we intended it, those preprocesses are not in a conflict with each other to our best knowledge.
Thank you for your interest in our work!