Attribute Discovery Dataset

Our Attribute Discovery Dataset consists of 4 broad shopping categories (bags, earrings, ties, and shoes). The data set was collected from like.com and includes both images and associated textual descriptions. It was collected for the system described in the paper: "Automatic Attribute Discovery and Characterization from Noisy Web Images".

Download (403 MB):
Attribute Discovery Dataset

To unpack the files into your current directory under unix use the command:
tar zxvf attributedata.tar.gz

The downloaded directory contains images and text files for 4 broad categories (bags_*, earrings_*, ties_*, womens_*), where * represents a number of subcategories, e.g. bags_shoulder contains images and descriptions for shoulder bags, bags_evening contains images for evening bags, etc. For each image there are two files with the descriptive text (descr*.txt) and original url (orig_url*.txt) respectively.

If you have a very fast connection it may help to limit the download rate. For example:
wget --limit-rate=500k "http://tamaraberg.com/attributesDataset/attributedata.tar.gz".

Reference:

This dataset is for academic research purposes only. If you use our dataset please reference:

Automatic Attribute Discovery and Characterization from Noisy Web Images [pdf]
Tamara L. Berg, Alexander C. Berg, Jonathan Shih, European Conference on Computer Vision (ECCV), 2010