Itcan be seen, except the extreme case with very smallk (e.g.k = 1), the results are generally stable, further corroborating the robustness of ranking statistics. We also carry out experiments using "hard" and "soft" cosine similarity. For the "hard" cosine similarity, we simply adopt athreshold (0.9 inour experiments) onthe score toget binary pseudo labels. While for the "soft" cosine similarity, we directly take the score as soft pseudo labels. Wechoose tousesoftranking statistics because webelievethe continuous similarity better reflect the actually similarity of objects than the binary score. This is important for the pairs with a similarity score around 0.5, for which the binary score is not very reliable.

In this paper, we tackle the problem of novel visual category discovery, i.e., grouping unlabelled images from new classes into different semantic partitions by leveraging a labelled dataset that contains images from other different but relevant categories.