This dataset provides a plattform to benchmark transfer-learning algorithms, in particular attribute base classification [1,2]. It consists of 30475 images of 50 animals classes with six pre-extracted feature representations for each image. The animals classes are aligned with Osherson's classical class/attribute matrix [3,4], thereby providing 85 numeric attribute values for each class. Using the shared attributes, it is possible to transfer information between different classes.


Attribute based classification and the data is described in:

[1] C. H. Lampert, H. Nickisch, and S. Harmeling. "Learning To Detect Unseen Object Classes by Between-Class Attribute Transfer". In CVPR, 2009 (pdf)

[2] C. H. Lampert, H. Nickisch, and S. Harmeling. "Attribute-Based Classification for Zero-Shot Visual Object Categorization". IEEE T-PAMI, 2013 (pdf)

The class/attribute matrix was originally created by:

[3] D. N. Osherson, J. Stern, O. Wilkie, M. Stob, and E. E. Smith. "Default probability". Cognitive Science, 15(2), 1991.

[4] C. Kemp, J. B. Tenenbaum, T. L. Griffiths, T. Yamada, and N. Ueda. "Learning systems of concepts with an infinite relational model". In AAAI, 2006.


Version 1.0, May 13th 2009

Because the dataset is rather large, the download is split into separate archives:

  • Base package (1M) including the class/attribute table:AwA-base.tar.bz2 (everybody needs this)
  • Color Histogram features (124M): AwA-features-cq.tar.bz2
  • Local Self-Similarity features (30M): AwA-features-lss.tar.bz2
  • PyramidHOG (PHOG) features (28M): AwA-features-phog.tar.bz2
  • SIFT features (44M): AwA-features-sift.tar.bz2
  • colorSIFT features (44M): AwA-features-rgsift.tar.bz2
  • SURF features (49M): AwA-features-surf.tar.bz2
  • DECAF (fc7 layer of 7-layer CaffeNet, pretrained on ILSVRC2012) features (122M): AwA-features-decaf.tar.bz2
  • VGG19 (fc7 layer of very deep 19-layer CNN, pretrained on ILSVRC2014, stacked from 10 image crops) features (1.2G): AwA-features-vgg19.tar.bz2 (NEW!)
  • Source code (30K) illustrating DAP and IAP methods: AwA-code.tar.bz2
    Addendum: new attributes.py script that work with recent versions of Shogun
  • Full Image Set in JPEG format: not directly downloadable for copyright reasons
  • Results

    Attribute based classification results (class-averaged multiclass accuracy):

    • direct-attribute prediction (DAP): 40.5% [1]       41.4% [2]
    • indirect-attribute prediction (IAP): 27.8% [1]       42.2% [2]
    • Attribute/Hierchical Label Embedding (AHLE): 43.5%;   [5]
    • Archive with confusion matrices and precision-recall curves AwA-results.tar

    Unpublished results (latest update: 21/07/2015):

    methodfeatures/kernelmulticlass accmean class AUCmean attributes AUCcomments
    DAPDECAF (d=4096)/linear47.7% 83.1 ± 8.30.73 ± 0.14default split
    DAPVGG19 (d=40960)/linear57.5% 86.3 ± 9.50.78 ± 0.14default split

    more results to follow...

    Other Publications using the Dataset (list not up-to-date)

    [5] Z. Akata, F. Perronnin, Z. Harchaoui, C. Schmid. "Label-embedding for attribute-based classification". CVPR 2013 (pdf)