NUS-WIDE (Official Repo)

Dataset Access

The dataset can be accessed from huggingface.

Dataset Overview

The NUS-WIDE dataset is a large-scale multi-label image dataset that can be widely used for image classification and multi-label learning tasks. It contains 269,648 images collected from Flickr, annotated with 81 concept labels .

Training set：161,789 images
Test set：107,859 images

Dataset Directory Structure

NUS-WIDE
├── Groundtruth
│   ├── TrainTestLabels
│   │   ├── Labels_zebra_Train.txt 
│   │   ├── Labels_zebra_Test.txt
│   │   ├── Labels_window_Train.txt
│   │   ├── Labels_window_Test.txt
│   │   └── ···
│   ├── AllLabels
│   │   ├── Labels_zebra.txt
│   │   ├── Labels_window.txt
│   │   └── ···
├── ConceptsList
│   ├── Concepts81.txt
│   └── Concepts81.mat

NUS_WID_Tags
├── Train_Tags81.txt
├── Train_Tags1k.dat
├── Test_Tags81.txt
├── Test_Tags1k.dat
├── TagList1k.txt
├── Final_Tag_List.txt
├── All_Tags.txt
├── AllTags81.txt
├── AllTags1k.txt

After downloading the dataset, it contains two zip files: NUS-WIDE.zipand NUS_WID_Tags.zip.

Folder and File Description

NUS-WIDE.zip

This archive contains two folders: Groundtruth and ConceptsList.

- Groundtruth

This folder contains two subfolders: TrainTestLabels and AllLabels.

• TrainTestLabels: Contains 162 .txt files in total.
  • Labels_<concept>_Train.txt and Labels_<concept>_Test.txt represent the labels for 81 concepts on the training and test sets, respectively.
  • Each file contains one row per image, with a binary value (0 or 1) indicating whether the image belongs to the corresponding concept.
  • The training set contains 161,789 samples, and the test set contains 107,859 samples.
• AllLabels: Contains 81 .txt files, each corresponding to one concept.
  • Each file contains 269,648 lines, with 0 or 1 indicating whether the image belongs to that concept, covering the full dataset (training + test).

- ConceptsList

This folder provides the list of 81 predefined concept names used in the dataset.

NUS_WID_Tags.zip

This archive contains 9 files, with descriptions as follows:

- Train_Tags81.txt, Test_Tags81.txt

Contain the labels for the training and test sets in terms of the 81 predefined concepts.

• Each row is an 81-dimensional binary vector, where 1 indicates that the sample belongs to the corresponding concept, and 0 otherwise.

- Train_Tags1k.dat, Test_Tags1k.dat

Contain the labels for the training and test sets in terms of 1000 tags.

• Each row is a 1000-dimensional binary vector representing tag presence.

- TagList1k.txt

A list of the 1000 tag names used in the 1k-tag representation.

- Final_Tag_List.txt

A refined list of 5018 tags, as referenced in the original NUS-WIDE paper:

NUS-WIDE: A Real-World Web Image Database from National University of Singapore

These tags were filtered and processed from the original Flickr tag data.

- All_Tags.txt, AllTags81.txt, AllTags1k.txt

These files provide binary matrices indicating tag presence for all 269,648 images:

• All_Tags.txt: 269,648 × 5018 binary matrix (5018 tag space)
• AllTags81.txt: 269,648 × 81 binary matrix (concept label space)
• AllTags1k.txt: 269,648 × 1000 binary matrix (1000 tag space)

Citation

If you use the NUS-WIDE dataset in your research or publications, please cite the following paper:

@inproceedings{chua2009nus,
  title={Nus-wide: a real-world web image database from national university of singapore},
  author={Chua, Tat-Seng and Tang, Jinhui and Hong, Richang and Li, Haojie and Luo, Zhiping and Zheng, Yantao},
  booktitle={Proceedings of the ACM international conference on image and video retrieval},
  pages={1--9},
  year={2009}
}