Implementing algorithm from paper by Zhang , et al

Paper: http://crcv.ucf.edu/papers/cvpr2016/Zhang_CVPR2016.pdf
NUS-WIDE Dataset: http://lms.comp.nus.edu.sg/research/NUS-WIDE.htm

/Dataset contains compressed files for constructing dataframes for the NUS-WIDE dataset. For the raw data, download directly from the link above.

point /Dataset/downloader.py to the appropriate file containing img urls to download images.

There are a large number of broken links in NUS-WIDE-urls.text if it is crucial to have access to all 260k images, request access from the authors of the dataset at the NUS-WIDE Dataset website.

• Concepts are the image classes. There is some overlap in the classes corresponding to fine-grained image classification (for example there exists animal and bird classes).
• Groundtruth are solutions to both training and test sets for assigning concept to image.
• Tags are more general categorical assignments with a high degree of overlap between tags.