Time and Space-Efficient Template Based Indexing

点击翻译按钮获取中文摘要

Abstract IIS-9907142Hebert, MartialCarnegie Mellon University$93,733 - 12 mos.Time and Space-Efficient Template Based IndexingThis is the first year funding of a three year continuing award. Under this award, the PI will develop an approach to image-based object recognition which directly addressesthe fundamental challenges posed by the object recognition problem: sensitivity to changes in pose and illumination, computational complexity with respect to the number and complexity of the models, and potentially prohibitive storage requirement for representing the models.The approach is based on computing the responses ofappropriate filters in images of the model objects and of the observed scenes. The research conducted under this grant involves the mathematical modelingof filters that have desirable properties with respect to changes in pose and illumination. Amongst the possible classes of filter, the class of steerable filters is usedas a starting point for this analysis, owing to their properties with respect to groups of image transformations. These properties are exploited to develop an efficient indexingscheme for recognition. The main contribution of this research is an efficient and robust approach to matching and indexing which may be used directly for recognition, and also may be used in conjunction with other recognition techniques, such as feature-based techniques. In addition, the research will contribute to a computational theory of steerable filters which will convert the mathematical properties of the filters to algorithms relevant to a broad range of recognition, image matching, and image analysis problems.