NUS-WIDE
A Real-World Web Image Dataset from National University of Singapore
Here we introduce a web image dataset created by Lab for Media Search in National University of Singapore. The dataset includes: (1) 269,648 images and the associated tags from Flickr, with a total number of 5,018 unique tags; (2) six types of low-level features extracted from these images, including 64-D color histogram, 144-D color correlogram, 73-D edge direction histogram, 128-D wavelet texture, 225-D block-wise color moments and 500-D bag of words based on SIFT descriptions; and (3) ground-truth for 81 concepts that can be used for evaluation. Based on this dataset, we identify several research issues on web image annotation and retrieval. We also provide the baseline results for web image annotation by learning from the tags using the traditional k-NN algorithm. The benchmark results show that it is possible to learn models from these data to help general image retrieval.
Download
The dataset, including low-level features, groundtruth, tags, concept list, image list and original urls, can be obtained through following links. The exif and geo-info for the images from a fraction of the dataset are available here. For more detailed descriptions of the dataset, see specification below.
- 1. Low-Level Features (1.1GB)
- 2. Groundtruth (1.1MB)
- 3. Tags (25MB)
- 4. Concept List (448KB)
- 5. Image List (4.6MB)
- 6. Image Urls (8.7MB)
We also provide (1) a light version of NUS-WIDE dataset, named NUS-WIDE-LITE, (2) one object image dataset, named NUS-WIDE-OBJECT and (3) one scene image dataset, named NUS-WIDE-SCENE. The specifications of these three subsets are detailed in the paper.
- 1. NUS-WIDE-LITE (110MB)
- 2. NUS-WIDE-OBJECT (62MB)
- 3. NUS-WIDE-SCENE (70MB)
For any question regarding the dataset, please contact Mr. Xindi Shang (shangxin AT comp.nus.edu.sg). The raw images can be obtained via sending a request email to us. Specifically, the researchers interested in the dataset should download and fill up the Agreement and Disclaimer Form and send it back to us. We will then email you the instructions to download the images at our discretion.
Citation
Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yan-Tao Zheng. "NUS-WIDE: A Real-World Web Image Database from National University of Singapore", ACM International Conference on Image and Video Retrieval. Greece. Jul. 8-10, 2009. [pdf] [Bibitex entry]
Specification
Low-Level Features
Each row in every low-level feature file corresponds to the image in the corresponding row of Imagelist.
- Normalized_CH : 64-D color histogram with each row represents an image.
- Normalized_CORR : 144-D color correlogram with each row represents an image.
- Normalized_EDH : 75-D edge direction histogram with each row represents an image.
- Normalized_WT : 128-D wavelet texture with each row represents an image.
- Normalized_CM55 : 225-D block-wise color moments with each row represent an image.
- BoW_int : 500-D bag of words with each row represent an image.
We separate the dataset into two parts, the first part contains 161,789 images for training and the second part contains 107,859 images for evaluation. So each low-level feature file is separated into two files which include 161,789 and 107,859 rows of features respectively.
- Normalized_CH = Train_Normalized_CH + Test_Normalized_CH
- Normalized_CORR = Train_Normalized_CORR + Test_Normalized_CORR
- Normalized_EDH = Train_Normalized_EDH + Test_Normalized_EDH
- Normalized_WT = Train_Normalized_WT + Test_Normalized_WT
- Normalized_CM55 = Train_Normalized_CM55 + Test_Normalized_CM55
- BoW_int = BoW_Train_int + BoW_Test_int
Groundtruth
AllLabels : this fold includes 81 files corresponding to 81 concepts respectively. For a certain concept, every row in the file corresponds to the groundtruth of the image in the corresponding row of Imagelist.txt. TrainTestLabels : every groundtruth file is separated into two parts according to the separation of the dataset. Thus this fold includes 162 files corresponding to the training and testing groundtruth of the 81 concepts.
Tags
AllTags81 : this file includes the labels extracted from the tags associated to the images for the 81 concepts. Each column represents a concept and each row includes the 81 labels extracted from the associated tags for the corresponding images. AllTags81(i, j) = 1 means the associated tags of the i-th image include the j-th concept in Concepts81.txt. Otherwise, AllTags81(i, j) = 0 means the associated tags of the i-th image does not include the j-th concept in Concepts81.txt.
- AllTags81 = Train_Tags81 + Test_Tags81
Train_Tags81, Test_Tags81 : the file AllTags81 is separated into two parts according to the separation of the dataset. AllTags1k : this file includes the labels extracted from the tags associated to the images for the top 1,000 concepts in the Fianl_Tag_List.
- AllTags1k = Train_Tags1k + Test_Tags1k
Train_Tags1k, Test_Tags1k : the file AllTags1k is separated into two parts according to the separation of the dataset. AllTags : each row includes the raw tags crawled from www.flickr.com for the image in the corresponding row of Imagelist.txt. Fianl_Tag_List : the tag list includes 5,018 tags extracted from the associated tags of the dataset. The tags are sorted according to their frequencies.
Concept List
Concepts81 : this file includes the 81 concepts in alphabetical order.
Image List
Imagelist : the list of raw images extracted from www.flickr.com
- Imagelist = TrainImagelist + TestImagelist
TrainImagelist : the list of images for training.
TestImagelist : the list of images for testing.