Intrinsic generalization analysis of low dimensional representations

Abstract

Low dimensional representations of images impose equivalence relations in the image space; the induced equivalence class of an image is named as its intrinsic generalization. The intrinsic generalization of a representation provides a novel way to measure its generalization and leads to more fundamental insights than the commonly used recognition performance, which is heavily influenced by the choice of training and test data. We demonstrate the limitations of linear subspace representations by sampling their intrinsic generalization, and propose a nonlinear representation that overcomes these limitations. The proposed representation projects images nonlinearly into the marginal densities of their filter responses, followed by linear projections of the marginals. We use experiments on large datasets to show that the representations that have better intrinsic generalization also lead to better recognition performance. q

Face Detection Using Spectral Histograms and SVMs

“…While most representations used in face detection are justified only based on empirical results, the sufficiency of the spectral histogram representation is shown systematically through statistical sampling. As our representation is generic in nature, it can be easily adapted to other forms of object detection and other tasks such as face recognition [16], object recognition [16], and texture classification [18]. With these results, we expect the spectral histogram representation may provide a unified representation for effective object detection and recognition.…”

“…For example, our approach is generic in nature and is applicable to other forms of object detection and recognition, not solely face detection. As shown by Liu et al [16,15,14], the spectral representation can be used to enhance the performance on such tasks as face and object recognition. This could lead to the possibility of integrating the two systems as one.…”

“…By analyzing generalization performance, Liu et al [16] showed that a necessary condition for a good representation is that it should group only images from the same class together. One way to analyze the intrinsic generalization of a given representation is to generate random samples from the set consisting of images with identical representations [16]. Technically, given a representation G and an image I obs from a known class, we want to sample from {I|G(I) = G(I obs )}.…”