I have found no similar questions to this online, or answers for that matter.
I am using cameras that output a grayscale image, which I feed into a Yolov8 object detection model (Specifically yolov8m-pose). What I have been doing during training and inference is converting the 1-channel data to 3-channel, which of course just copies the data to every channel, and then using it like you would for any RGB image.
I felt like there must have been a more optimal way to go around this, so I decided to adjust the Yolov8 pipeline to train and do inference with 1-channel images. This will decrease the input layer count threefold, but the rest of the model remains the same size (When still using the yolov8m-pose model) and thus there is no real performance gain.
Now comes my actual question... The performance hardly gets any better. But when using the same model parameters as before, the core model is way larger compared to the amount of input layers. Is there an accuracy benefit to be obtained by training my model this way, compared to sticking with the 3-channel model? Can I, for instance, reduce the width and depth of the 1-channel model and still reach the same accuracy as the original 3-channel model?
I understand it is usually the case when you give up color, you reduce the amount of data your model uses. But in this use case there isn't any color data to begin with.