Changine image mode produce wrong results for resnet 50 model

While quantizing the resnet50-v1 model with image mode 0to1, for the example cat image, it predicts the class 285. For the image mode = neg1to1, it predicts class id: 281, but for image mode = 0to255, it predicts 416 which is wrong.
The command I am using for quantization is,
./image-classifier cat_285.png -image-mode=neg1to1 -model=resnet50-v1-7.onnx -model-input-name=data -dump-profile=profile.yml

Why just changing the image mode produces the wrong results?

Image mode changes how we normalize the input image data. If the model is trained with some specific normalization then we should try to use that same normalization to match to achieve the best accuracy.