Dynamic detection of visual entities

This paper tackles the issue of retrieving different instances of an object of interest within a given video document or in a video database. The principle consists of considering a semi-global image representation based on an over-segmentation of video frames. An aggregation mechanism is then applied in order to group a set of segments into an object similar to the query, under a global similarity criterion. We test the effectiveness of three different aggregation strategies, two of them based on a greedy approach, and the third one involving simulated annealing optimization. Experimental results on different color spaces show promising performances, with First Tier and Bull Eye detection rates of up to 70% and 88%, respectively. The integration of the method in a web-based video navigation system, allowing fast video object retrieval, is finally described.