Trying to implement a Keras CNN model on Unity but my outputs are always in the wrong range/scale.

[autumn tusk]

I'm trying to implement an Image Regression Keras CNN model on Unity. The input of the model must be 1080p images that have been downscaled by 4, so the form of the input should be (1, 270, 480, 3), at least it looks like that in Python. The CNN model takes in this singular frame and predicts 2 numerical values which are supposed to be x and y coordinates, in the downscaled range. The input must be taken from the perspective of the playable character, so it will be taken as screenshots from the camera. I tried to implement this model on Unity by converting it into an ONNX model and then using Barracuda. Because the pixel color values in Python and Unity are different I scaled all the pixels of the input frames by 255, and with this configuration I obtained outputs but all of the x-y coordinate outputs were capped at 255, which is wrong. I did the same thing with the ONNX-runtime method but I ended up with the same outputs. Are there any other ways of implementing the CNN model on Unity? Or would it be possible to solve the problems with the methods I tried?