A Measure of Perceptual Aliasing in Image Descriptors

Abstract

This thesis is concerned with a measure of perceptual aliasing in image descriptors. Perceptual aliasing occurs when the one-to-one mapping relations between world states (objects) and their representation (descriptors) are not maintained. Our method measures the discriminating power of an image descriptor in terms of its ability to distinguish between images of different objects and to match images of the same object. Specifically, our method runs spectral clustering on the similarity matrix computed with descriptors of known image clusters and measures the performance of an image descriptor by its ability to maintain the original clusters, using two indices, MRI-1 and MRI-2, that are based on the Rand index. Experiments on MRI versus precision and recall show that our proposed metrics are more appropriate for applications such as content-based image retrieval in which image clustering is a critical step.