How to filter the original dataset into a subone used in the paper?

[qinnzou]

The process contains two steps:
Firstly, we calculate the number of images contained in the nuswide, that is 195,834. Then, we bulid a subset of images associated with the 21 most common labels, and each label associates with at least 5000 images. Under this setting, there are 195,834 images left.

Second, we compare the concept lists of nuswide-21 and coco , getting 4 empty classes——animal, person, road, vehicle. These four classes are filtered and the data that contain these empty labels are removed from the subset obtained in the first step simultaneously. At this point, there are 100,303 images left.