The fact that our system performance noticeably degraded means the number of new birds has become a significant percentage of the total number of real-world birds, and the fact that our system is not doing well, is probably because it suffers from high bias as real-world data shifted (i.e. training our system on detecting black cats and expecting it to do well on other cats), to say that the first thing to try is not to look for ways to include the pictures of the new species in our training dataset sounds counterintuitive to me, please let me know why this isn’t the case?