August 23, 2000
Barus & Holley, Rm. 312 - Eng. bldg. 3rd floor, 3pm.
Xavier Orriols from Barcelona is visiting us over the summer and working on some vision problems relevant to our Digital Archaeology project. Xavier will be at the SHAPE lab until the end of August.
The object recognition process via internal representations (abstract models) needs a measure of similarity to compare the responses in visual field with the learned object model. Such measure has to relate if it exists a stimulus equivalence between the responses in the image and the produced one by the stored model. Unlike bottom-up techniques, this process is a high-level and goal-driven task, and considers the context of the visual attention environment.
The target object-class detection problem affected by visual appearance variability needs, for any object-class, a membership function that is able to identify the degrees of freedom (intrinsic dimensionality) of the category in the internal representation. Such a measure is necessary to build a saliency map in order to stand out the regions of interest, i.e. candidate image co-ordinate points that can allocate the searched structure. The model of internal representation object categories into a probabilistic framework defines the approach of target-location problem in terms of maximum likelihood (ML) estimation.