The idea is that juvenile annotations are so difficult that it may possibly cause the model to make more false positives trying to find juveniles. Juvenile salmon aren't necessary to be counted, so it may be better to ignore them if it improves the performance.