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.