ImageCLEF 2012 Image annotation and retrieval dataset (MIRFLICKR)

DESCRIPTION For this task, we use a subset of the MIRFLICKR (http://mirflickr.liacs.nl) collection. The entire collection contains 1 million images from the social photo sharing website Flickr and was formed by downloading up to a thousand photos per day that were deemed to be the most interesting according to Flickr. All photos in this collection were released by their users under a Creative Commons license, allowing them to be freely used for research purposes. Of the entire collection, 25 thousand images were manually annotated with a limited number of concepts and many of these annotations have been further refined and expanded over the lifetime of the ImageCLEF photo annotation task. This year we used crowd sourcing to annotate all of these 25 thousand images with the concepts. On this page we provide you with more information about the textual features, visual features and concept features we supply with each image in the collection we use for this year's task. TEXTUAL FEATURES All images are accompanied by the following textual features: - Flickr user tags These are the tags that the users assigned to the photos their uploaded to Flickr. The 'raw' tags are the original tags, while the 'clean' tags are those collapsed to lowercase and condensed to removed spaces. - EXIF metadata If available, the EXIF metadata contains information about the camera that took the photo and the parameters used. The 'raw' exif is the original camera data, while the 'clean' exif reduces the verbosity. - User information and Creative Commons license information This contains information about the user that took the photo and the license associated with it. VISUAL FEATURES Over the previous years of the photo annotation task we noticed that often the same types of visual features are used by the participants, in particular features based on interest points and bag-of-words are popular. To assist you we have extracted several features for you that you may want to use, so you can focus on the concept detection instead. We additionally give you some pointers to easy to use toolkits that will help you extract other features or the same features but with different default settings. - SIFT, C-SIFT, RGB-SIFT, OPPONENT-SIFT We used the ISIS Color Descriptors (http://www.colordescriptors.com) toolkit to extract these descriptors. This package provides you with many different types of features based on interest points, mostly using SIFT. It furthermore assists you with building codebooks for bag-of-words. The toolkit is available for Windows, Linux and Mac OS X. - SURF We used the OpenSURF (http://www.chrisevansdev.com/computer-vision-opensurf.html) toolkit to extract this descriptor. The open source code is available in C++, C#, Java and many more languages. - TOP-SURF We used the TOP-SURF (http://press.liacs.nl/researchdownloads/topsurf) toolkit to extract this descriptor, which represents images with SURF-based bag-of-words. The website provides codebooks of several different sizes that were created using a combination of images from the MIR-FLICKR collection and from the internet. The toolkit also offers the ability to create custom codebooks from your own image collection. The code is open source, written in C++ and available for Windows, Linux and Mac OS X. - GIST We used the LabelMe (http://labelme.csail.mit.edu) toolkit to extract this descriptor. The MATLAB-based library offers a comprehensive set of tools for annotating images. For the interest point-based features above we used a Fast Hessian-based technique to detect the interest points in each image. This detector is built into the OpenSURF library. In comparison with the Hessian-Laplace technique built into the ColorDescriptors toolkit it detects fewer points, resulting in a considerably reduced memory footprint. We therefore also provide you with the interest point locations in each image that the Fast Hessian-based technique detected, so when you would like to recalculate some features you can use them as a starting point for the extraction. The ColorDescriptors toolkit for instance accepts these locations as a separate parameter. Please go to http://www.imageclef.org/2012/photo-flickr/descriptors for more information on the file format of the visual features and how you can extract them yourself if you want to change the default settings. CONCEPT FEATURES We have solicited the help of workers on the Amazon Mechanical Turk platform to perform the concept annotation for us. To ensure a high standard of annotation we used the CrowdFlower platform that acts as a quality control layer by removing the judgments of workers that fail to annotate properly. We reused several concepts of last year's task and for most of these we annotated the remaining photos of the MIRFLICKR-25K collection that had not yet been used before in the previous task; for some concepts we reannotated all 25,000 images to boost their quality. For the new concepts we naturally had to annotate all of the images. - Concepts For each concept we indicate in which images it is present. The 'raw' concepts contain the judgments of all annotators for each image, where a '1' means an annotator indicated the concept was present whereas a '0' means the concept was not present, while the 'clean' concepts only contain the images for which the majority of annotators indicated the concept was present. Some images in the raw data for which we reused last year's annotations only have one judgment for a concept, whereas the other images have between three and five judgments; the single judgment does not mean only one annotator looked at it, as it is the result of a majority vote amongst last year's annotators. - Annotations For each image we indicate which concepts are present, so this is the reverse version of the data above. The 'raw' annotations contain the average agreement of the annotators on the presence of each concept, while the 'clean' annotations only include those for which there was a majority agreement amongst the annotators. You will notice that the annotations are not perfect. Especially when the concepts are more subjective or abstract, the annotators tend to disagree more with each other. The raw versions of the concept annotations should help you get an understanding of the exact judgments given by the annotators