Image retreval based on location and time

31.05.2006

Abstract:
Events are a high-level concept based on geographical location, time and date that has been used to manage image collections, because people think in terms of events when they organize and search their personal photo albums, whether the collection is digital or not. Existing systems for managing personal images use event segmentation as a method of automatically organizing different events for more effective image retrieval.

However, in case of a public image collection, the event segmentation process becomes more difficult, because the images are possibly contributed by multiple sources, which requires that both temporal and spatial information is used to segment events. This is in contrast to a personal image collection where the image contributors are often just one, and the different events of the collection can be ordered successively and coherent in time and location. In this case, temporal information is enough to separate and organize different events.

The work of this thesis consisted of proposing a temporal and spatial image retrieval method for use with an image collection with multiple image contributors. The images of the collection are connected to events. Different events are separated at a fine granularity, organized by location, and enabled for retrieval.

A prototype is implemented based on the proposed method, and focus on the user aspects of image retrieval; matching the query to images of events and presenting them successively in time.

Difficulties in the detection of different events in a public image collection suggest that a segmentation method based on temporal and spatial information has limitations; the segmentation method divides events into several fragments that are difficult to interconnect by using only temporal and spatial information.