One of the key problems in appearance-based vision is understanding how to use a set of labeled images to classify new images. Classification systems that can model human performance often make use of similarity judgements that are non-metric. Existing condensing techniques for finding class representatives which are ill-suited to deal with non-metric dataspaces are presented. The importance of ideas learned from the experience of improving performance using both synthetic and real images are discussed.
|Number of pages
|Published - 1998
|Proceedings of the 1998 IEEE 6th International Conference on Computer Vision - Bombay, India
Duration: 4 Jan 1998 → 7 Jan 1998
|Proceedings of the 1998 IEEE 6th International Conference on Computer Vision
|4/01/98 → 7/01/98