A robust distance measure for the retrieval of video objects

We propose a new method for measuring the similarity between two arbitrarily shaped video objects. Our method is based on comparing the low-level still features of the representative planes of video objects. We demonstrate the performance of the proposed method in a shape retrieval system in which boundary- and region-based still shape features were employed to retrieve video objects. The experimental results show that i) the retrieval performance using the proposed similarity matching technique significantly outperforms the commonly used feature vector averaging technique and ii) the distance measure performs robustly when the content of the object changes in time