Visual-search

Final version of CDVS evaluation software, produced at 97th MPEG meeting can be found here.

The final version of the Paris Dataset, with annotations and images is available here as a zip archive: Paris Database

The CDVS Evaluation Software is available here as a zip archive.

The SVN repository for the CDVS software will be available soon.

We collected approximately 1100 images, 300 video clips for 200 buildings on campuses of Hong Kong Polytechnic University and Northwestern University and surrounding areas. In addition, 400 thumbnail images of 640×480 are also generated for the video clips.
The annotation consists of 4 parts:

Hong Kong Data Set Icons

Northwestern Data Set Icons

1. matching_pairs: 400 entries
2. non_matching_pairs: 6000 entries
3. non_matching_pairs_12000: 12000 entries
4. retrieval: 600 cases

The data set is being checked and once ready will be uploaded.

These are the cross-checked annotation files of the SudParis DB. The zip-file contains three text files: xcheck_annotations_sudparis.txt (annotation file for the retrieval task with the format ), xcheck_match_pairs_sudparis.txt (list of matching pairs), and xcheck_non_matching_sudparis.txt (list of non-matching image pairs).

The duplicated pictures were removed from the annotations and it was proved that the interest objects (e.g. Eiffel Tower, Arc de Triomphe, etc.) do not appear in both images of a non-matching pair.

Some issues still need to be dealt with. As Sang-il mentioned, there are some objects which appear repeated in several images of the dataset (e.g. streetlights, traffic signs, flags, even cars or trees), which might lead to misleading matches or object localisation, if the objects were too big. Any suggestion will be welcome.

The cross-checked annotation files can be downloaded from the following link:

xcheck_sudparis

PS Some file names include a white space, which was removed in the annotation files. The file names should be changed accordingly.

This contribution contains nearly five hundreds pictures about buildings and landscapes in Paris which are captured mainly by traditional cameras and mobile phones, from different distances and perspective, with different weather and illumination conditions so that their resolution and quality are very divers. For the time being, our database has 35 categories in total; and each category represents one object which is either a building or a landscape in Paris.
Every picture is accompanied with one annotation file which indicates the type of picture (can be either “reference” or “query”) and represents objects of interest that it contains

Example 1:

This photo belongs to the category “Arc de Triomphe” which includes 23 images in JPG format

Example 2:

This is an example of a photo can contain more than one object: “Saint-Etienne du Mont” and “Pantheon Monument”. So the annotation file of the photo will look like as below:
query [2, paris_34, paris_28]
Where

• query: will be used for tests
• paris: database name
• 34, 28: ID number of “Saint-Etienne du Mont” and “Pantheon Monument” respectively

You can download this dataset here
A completed version of the document can be also found here

The goal of these pages is to put together the dataset and other information related to visual serch project of MPEG