I really don't know where else to ask this.

I'm training an object detection model to recognize bad paint splotches like this. However, the only training data I have access to has been corrupted by idiots in another department who overwrote the original images and coloured red circles all over them.

During training, the filter passes over the image. Obviously the red label around #3 is a shitshow and can't be used in training or testing. However, if I label #1 and draw my object detection training box within the red such that it doesn't touch the red, will it still the model start looking for items within red pen marks as the kernel slides over it? Finally, if I just label items like #2 and ignore items like #1, will the model be punished for identifying #1 but I haven't labelled it?

Any thoughts are appreciated.