Improving computer vision for shape matching
There are two approaches that are normally followed for computer vision tasks, a bottom-up or a top-down. Yet, both approaches involve "gray areas" as details are often overlooked in the bottom-up practice, while the top-down methodologies require a lot of computation. Answering this need the DSSCV project focused on the simultaneous solution of computer vision tasks at all levels with the aid of the deep scale-space structure of images. Thereby, the use of a multi-scale singularity structure of images may solve computer vision tasks in the most elegant, robust and efficient way. More specifically, the project resulted in the development of theory and practice in singularity theory, scale-space theory and algorithmics for deriving efficient algorithms to address computer vision tasks. Key areas of application of these algorithms include image databases, medical imaging, and image coding. One of the developed and implemented algorithms has been evaluated in terms of shape matching on the basis of a pre-symmetry diagram. The performance evaluation was conducted using a database of ninety-nine shapes that were divided in eleven classes. The current form of the matching algorithm was found unable only when addressing severe occlusion issues. Despite that, it was shown that the symmetry sets may be useful in shape characterisation and very effective in shape retrieval. Hence, this technology may be employed for querying in content-based image databases. For further information click at: http://www1.itu.dk/sw1953.asp(opens in new window)
