Exemplar Based Recognition of Visual Shapes

In the talk I will presents an approach of visual shape recognition based onexemplars of attributed keypoints. The work may be seen as a refinementof a previous work [1]. The work is aimed at object shape categorizationby image semantic report. Training is performed by storing exemplarsof keypoints detected in labeled training images. Recognition is made bykeypoint matching and voting according to the labels for the matchedkeypoints. The matching is insensitive to rotations, limited scalings andsmall deformations. The recognition is robust to noise, background clutterand partial occlusion. Recognition is possible from few training imagesand improve with the number of training images. The simple trainingprocedure, eliminating off-line shape prototype definitions, facilitates theuse when many shapes are to be learned. One feature of the chosenkeypoint descriptor is that an image of the recognized structures can besynthesized based on the matched keypoints. References[1] S.I. Olsen: End-Stop Exemplar based Recognition, Proceedings of the 13thScandinavian Conference on Image Analysis, 2003, pp. 43-50. Also availablefrom ftp://ftp.diku.dk/diku/users/ingvor/SHAPE03/