Is there a "best practice" for using networks with color input (like gaze-estimation-adas-0002) with IR illuminated grayscale cameras as in the OAK D Pro?

Retraining on "IR images" seems like a potential last resort. Manipulating the grayscale image could be an easier solution.

Hi @dexter
If the trained model is not channel feature sensitive, you can just convert the grayscale image to RGB by stacking three grayscale layers together.

Snippet from https://docs.luxonis.com/projects/api/en/latest/samples/MobileNet/mono_mobilenet/:

# Convert the grayscale frame into the nn-acceptable form
manip.initialConfig.setResize(300, 300)
# The NN model expects BGR input. By default ImageManip output type would be same as input (gray in this case)
manip.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p)

Thanks,
Jaka