News and events
2 October 2013
Classification Regardless of Depictive Style
Presented By Dr Peter Hall, Reader, University of Bath
- N.1.12, Haslegrave Building
About this event
Abstract: Classification is one of the most active yet open areas in all of Computer Vision. In some regards the field is exceptionally well developed: it is possible for a machine to learn possibly thousands of object classes - that is nouns such as car, bike, dog, elephant, such that never before seen examples can be recognised. What is more, recognition can be robust to large variations in point of view, lighting, occlusion and other common problems that complicate the issue.
Amongst all this work, the problem of classification regardless of depictive style is under researched, so that nearly all algorithms fail to classify objects in paintings and drawings; for example, nearly all algorithms tacitly assume and are tuned to photographic input.
This talk will describe research that shows how to build a classifier that is robust to depiction. Using several images from class as source, our classifier builds a model to represent each class (one model per class). Each model captures structure and qualitative descriptions relating to its class. Empirical evidence shows we are able to classify across depictions at state of the art rates: we are about 10% to 50% more accurate. We will also talk about the uses of such a classifier to Computer Graphics. In particular the automatic production of art from photographs, in forms that are not possible otherwise.