Image Retrieval Showcase
Contact: Allan Hanbury, Vienna University of
Technology, Austria.
Database
An extended version of the IAPR-TC12 dataset of
20,000 vacation images. The version of the dataset used in the ImageCLEF 2006
campaign can be obtained on the IAPR TC12
Benchmark page. This dataset is a good simulation of a collection of photos
that could be stored on anybody's harddisk. The version of the dataset used in
the showcase event will be modified as follows:
- New images will be added.
- The Description field will be removed from the image
annotations (all other fields will remain).
- Only the English annotations will remain.
- A few fields for some images will be left
blank, to simulate the poor annotation which is usually done for personal photo
collections.
This dataset for use in the image retrieval showcase
will be made available in the middle of April.
Types of Queries
For an idea of the types of queries, see
the queries used in the ImageCLEF 2006 campaign, also available on the IAPR TC12
Benchmark page. For the showcase event, two types of queries will be
provided:
- Text only queries: e.g. "Find images of snowy
mountains", "Find images of hotels with swimming pools", "Find images with three
people on a beach".
- Text queries with example images: e.g. "What
is the name of the monument in the provided image?" (i.e. find other images of
the same monument and look at the text annotation), "Find other images of Uncle
Fred (see attached image)".
Some example queries will also be made available in the middle of April.
Organization
Each participant should bring a laptop computer with their image retrieval system
and the dataset installed, and set it up at the event. Internet access will not be available. The queries will be
handed out one by one at the event. The "searcher" designated by each team will then use the system to find the
images from the dataset matching the query (note that the same searcher must do all the queries for a team).
The searches may be as interactive as desired and go through as many iterations as needed. The searches on
all the systems will happen in parallel.
Kudos will be awarded for the systems that:
- find the first correct image the fastest.
- find the most correct responses in the time limit.
- find the most correct images that other systems did not find.
- have the best looking user interface.
- find the most images of Uncle Fred.
- ...
During the second part of the event, queries on the systems will be tried out by a few members of the audience. The
most user-friendly system, easiest to learn system, best result display, ... will be chosen by the audience members.
None of the evaluation results will be published, the emphasis is on demonstrating the capabilities of the technology
for a well-defined task that interests many people: what is the best way to search my home photo collection?