search

bbenligiray / nus_wide_formatter_SRN

3 stars 0 forks source link

About nus_wide database #1

Open tianzhenz opened 5 years ago

tianzhenz commented 5 years ago

Hi, I have two questions about nus-wide database.

first question is that how to generate a nus_wide_test_label.txt ？such as 0000000100001......,Why are there multiple 1s and how to generate them？

The second is that what's relationship between cats and database？I could not find the folder of cats in the original dataset .Is the data set divided into 81 categories of cats,but I found that the original data set contains more than 700 folders .I am confused.

bbenligiray commented 5 years ago

1- In https://github.com/bbenligiray/nus_wide_formatter, I read the labels from the files provided with the dataset. However, those labels are not correct (no idea why). I got the txt files in this repo from https://github.com/zhufengx/SRN_multilabel (links in Prepare data under the readme). From what I can tell, those labels are correct, and give significantly better results. There are multiple 1s because this is a multi-label classification dataset.

2- The images are labeled in two different schemes: 1000 original Flickr tags (these are very noisy) 81 concepts, hand-labeled by the authors Only the 81 concepts are used in multi-label classification experiments, so they are the ones I've provided.