RGB Value Normalization

[pmiracle]

Hey,

Looking at the image classification application example, I couldn't find any part of the code that handles the normalization of pixel value from [0,255] to [-1,1] or [0,1] that I would expect from a CV model of that kind. 

Do you mind explaining how this implementation that is based on the mobilenet_v2_0 skips the need for normalization?

Thanks.

[Julian E.]

Hello @pmiracle,

In the example you linked, the deployed model is a model from model zoo: 

mobilenet_v2_0.35_128_fft_int8.tflite

This model takes a float32 input [-1,1] and contains the scale and offset:

The yaml from model zoo only indicate the scripts how to edit the getting started application for model zoo to deploy it on the stm32. ie:

If you run the deployment yaml throught the stm32ai_main.py, you can look at the edited and deployed application here:  /application_code/image_classification/STM32N6/

Have a good day,

Julian

[pmiracle]

Thanks for the reply,

This doesn't makes sense to me, because the camera on the DevKit creates RGB values in the range of 0-255 and there is no code in the example project that I have seen that translates these values to -1,1.

Any explanation?

[Julian E.]

Hello @pmiracle,

You're right, I misread. 

The float32[-1:1] is the output (look at the image).

The input is uint8[128,128,3] and most likely RGB values in the range of 0-255 as you pointed out.

Have a good day,

Julian

[pmiracle]

But MobileNet expects values [0,1], so what am I missing?

[Julian E.]

Hello @pmiracle,

There are two different cases depending on whether you’re talking about the float model or the quantized model.

Case 1:  Float (FP32 / FP16) MobileNetV2

• The original MobileNetV2 models trained on ImageNet usually expect input normalized to:
  • either [0,1] (just divide by 255)
  • or sometimes normalized further with mean/std (e.g. PyTorch convention: (x/255 - mean) / std).
• In TensorFlow Lite’s reference MobileNetV2 float models, the convention is simply [0,1].

So if you are using the non-quantized version, the comment “MobileNet expects values [0,1]” is correct. 

Case 2: Quantized (INT8 / UINT8) MobileNetV2

• When you quantize the model, the input is no longer float but integers.
• The quantization metadata (scale + zero_point) encodes how to map from uint8 [0,255] to the internal float domain.
• In your case: float = 0.0078431377 * (q - 127) which corresponds to an internal domain of [-1, +1].

  • q = 0 → float ≈ -1

  • q = 127 → float = 0

  • q = 255 → float ≈ +1

• But the TFLite/STM32 runtime expects you to feed uint8 [0,255], and it applies the scaling automatically.

So for quantized models, you do not normalize to [0,1] yourself. You just give raw 8-bit pixels, and the model’s quantization parameters handle the mapping.

Have a good day,

Julian