Compositional Dictionaries for Domain Adaptive Face Recognition

摘要: We present a dictionary learning approach to compensate for the transformation of faces due changes in view point, illumination, resolution, and so on. The key idea our is force domain-invariant sparse coding, i.e., designing consistent representation same face different domains. In this way, classifiers trained on codes source domain consisting frontal can be applied target (consisting poses, illumination conditions, on) without much loss recognition accuracy. first learn base dictionary, then describe each shift (identity, pose, illumination) using over dictionary. adapted expressed as linear combinations context recognition, with proposed compositional approach, image decomposed into representations given subject, illumination. This has three advantages. First, extracted subject across domains, enables pose insensitive recognition. Second, subsequently used estimate condition image. Last, by composing we also perform alignment normalization. Extensive experiments two public data sets are presented demonstrate effectiveness