search

PKU-ICST-MIPL / DRLIH_TMM2020

3 stars 3 forks source link

About nus-wide database #6

Open burhanbarakli opened 2 years ago

burhanbarakli commented 2 years ago

I think you are not running the nuswide and mrflickr databases for multilabel. Did you organize each category yourself? As an example, there are not enough images from every class for 21 classes in Nuswide (for example, there are 46 images from sunset), so you prepared similar images as a concept for 21 classes. If so, some of the images will need to be similar in the categories in 21 classes.

Can you share a ready dataset for nuswide and mrfliecker?

thanks.

magicfisk commented 2 years ago

We follow the setting of the [1], the subset of images associated with the 21 most frequent labels, where each label associates with at least 5,000 images. The method requires the triplet images, which are organized according to those 21 labels.

[1] Simultaneous Feature Learning and Hash Coding with Deep Neural Networks, CVPR 2015.