Profiset Download

The whole 20M collection consists of files with thumbnails of the original images, the image annotations, the DeCAF desctiptors, the five MPEG7 descriptors. Each image is uniquely identified by its integer ID that is present in each of the resources.

Link to the Profimedia

The web page on the Profimedia site for each image can be accessed using
http://www.profimedia.cz/identifier-of-the-image
Various sizes of the image as well as the title and keywords can be found on the web page.

Thumbnails

JPEG thumbnails of 130px were generated from the original images.
Download a ZIP file with all the thumbnails. The identifier of the image is the name of the thumbnail file (without the “.jpg” extension). Each image can be also accessed using a web service at
http://disa.fi.muni.cz/profimedia/images/identifier-of-the-image

Annotations

Each image is accompanied with a title and a set of keywords (30 in average).
Download a zipped CSV file with three columns: the image identifier, the title, and a list of keywords.

DeCAF Descriptors

We have extracted a DeCAF7 descriptor, which is a 4096-dimensional float vector, from each image.
Download a gzipped text file, where each record consists of two lines. First line contains a header with the image identifier
#objectKey messif.objects.keys.AbstractObjectKey identifier-of-the-image
The second line is a space-separated list of 4096 float numbers representing the descriptor.

Descriptors were extracted using the Caffe framework using the pre-trained model “BVLC Reference CaffeNet”. The layer “fc7″ of the deep neural network was used to get the descriptor.

To compute the distance (dissimilarity) of any two images, the Euclidean (L2) metric can be used. The MESSIF library (that is used in our demo) can be used to load and compute the distance using class messif.objects.impl.ObjectFloatVectorL2.

MPEG-7 Descriptors

We have also extracted a smaller global visual features as specified by MPEG-7 standard. In particular, we have extracted the Color Layout, Scalable Color, Color Structure, Edge Histogram, and Region Shape descriptors.
Download a gzipped text file, where each record consists of six lines.
The first line contains a header with the image identifier
#objectKey messif.objects.keys.AbstractObjectKey identifier-of-the-image
The second line represents the 16 bytes of YCbCr color space of the Color Layout descriptor (three comma-separated values of Y, Cr, and Cb semicolon-separated coefficients).
The third line is comma-separated list of 64 short integers that represents the Color Structure descriptor.
The fourth line is comma-separated list of 64 integers that represents the Scalable Color descriptor.
The fifth line is comma-separated list of 80 bytes numbers that represents the Edge Histogram descriptor.
The sixth line is comma-separated list of 36 bytes (linearized two-dimensional array of 12×3 angular and radial components) that represents the Region Shape descriptor.

To compute the distance (dissimilarity) of any two images, the combination of the distances of the five MPEG-7 descriptors can be used, as specified here. The MESSIF library (that is used in our demo) can be used to load and compute the distance using class messif.objects.impl.MetaObjectShapeAndColor, where the combination weights as well as the distance function Java implementations can be found.

Citation

If you will download and use the Profiset collection for research purposes, please, reference the following paper:

Budikova, P., Batko, M., and Zezula, P. (2011). Evaluation Platform for Content-based Image Retrieval Systems. In Proceedings of International Conference on Theory and Practice of Digital Libraries 2011, LNCS 6966, pages 130-142, Berlin: Springer. ISBN 978-3-642-24468-1.

If you use the DeCAF descriptors for efficiency evaluation, you can reference the following paper:

Novak, D., Batko, M., and Zezula, P. (2015). Large-scale Image Retrieval using Neural Net Descriptors. In Proceedings of SIGIR ’15, pages 1039-1040, ACM New York, NY, USA. ISBN 978-1-4503-3621-5.