nocaps

a large-scale benchmark for novel object captioning; the task of describing images containing visual concepts not seen in paired image-caption training data (Agrawal, Desai, Wang, Chen, Jain, Johnson, Batra, Parikh, Lee, Anderson)