A Location Annotation System for Personal Photos
Chufeng Chen, Michael Oakes and John Tait
University of Sunderland {chufeng.chen, michael.oakes, john.tait}@sunderland.ac.uk

Categories and Subject Descriptors: H.3.3
Information Search and Retrieval: Clustering and Search Process.

General Terms: Algorithms, Design, Human Factors. Keywords. Annotation, Clustering, Browsing. 1. INTRODUCTION
As the number of digital images in personal collections is increasing, it becomes necessary to support users of digital cameras with software tools for personal image management. We originally developed a time and location based clustering model to produce a browsing tool for collections of images stamped with GPS metadata [1]. The idea was to group the images into discrete events, where two images belonged to the same event if they were taken at similar times in nearby locations. We used the time clustering model of Platt et al. (2002) which first sorts the images in the collection by time, then marks the start of a new event whenever Log( gN )  K + 1 /(2d + 1) d Log( gN + i) , where gN+i is the

Agency (http://earth-info.nga.mil/gns/html/index.html). The key fields for us were the longitude, latitude, place name and geographic feature for each entry. The longitude and latitude stamps of each image were compared with the longitude and latitude of each entry in the gazetteer using Euclidian distance, and the closest location to where the photo was taken was found. The photo would then be annotated with the relevant place name such as "London Charing Cross" and the corresponding geographic feature "Railroad".

3. USER INTERFACE



i =-d

time gap between image i and image i+1, gN is the time gap between two successive images, d is the width of a sliding window of successive images (e.g. d =10). K is a suitable threshold (Platt et al. used K = log17). 1 /(2d + 1) d Log ( gN + i) is



i = -d

the average time gap between successive images in the window. Our location clustering model has two components: a latitude clustering algorithm and a longitude clustering algorithm. We used formulas which are analogous with that of Platt et al. (2000), except in that the images are initially sorted by latitude and longitude respectively rather than by time, and instead of gN and gN+1 being time differences, they are differences between adjacent images in latitude or longitude. Two images are considered to represent the same overall event if and only if they have been found to be in the same time event, the same latitude event, and the same longitude event. This system has been extended to allow users to retrieve images in response to text queries, which are matched against keyword annotations for each image, which may be either system or user-generated. The system consists of three stages: image metadata extraction, event clustering and image annotation by looking up keywords in a gazetteer, and matching of user text queries against the image annotations.

Figure 1. User interface The interface has two parts (see Figure 1): a command panel and a display panel. The command panel allows users to submit their queries. The display panel displays the search results, which are the 30 images whose annotations best match the query according to the Cosine Similarity Measure. For display, the images are clustered into discrete events, and each event is labeled with the relevant time and location. The individual images are also labeled with their individual similarity scores with respect to the query.

2. LOCATION NAME AND FEATURE ANNOTATION
The gazetteer we used was the UK file of the Geographic Names Data Base, maintained by the National Geospatial-Intelligence
Copyright is held by the author/owner(s). SIGIR'06, August 6­11, 2006, Seattle, Washington, USA. ACM 1-59593-369-7/06/0008.

4. REFERENCE
[1] Chen, C., Oakes, M. and Tait, J.: Browsing Personal Images Using Episodic Memory (Time + Location). Proceeding of ECIR, Vol. 1, (2006) 362-372

726