This paper presents a method for automatically learning 3-D object models from their appearance in 2D images. The appearance of an object is influenced by its shape, reflectance, pose, and illumination. Recognition is based on matching appearance rather than shape. A compact representation of object appearance is proposed, parametrized by pose and illumination. For each object, a large set of images is generated by varying pose and illumination. These images are compressed into a low-dimensional subspace, called the eigenspace, where the object is represented as a manifold. Given an unknown image, the system projects it into the eigenspace and recognizes the object based on its position on the manifold. Experiments show that with less than 20 eigenspace dimensions, recognition is accurate with an average pose error of about 1 degree. A near real-time recognition system with 20 complex objects has been developed. The paper discusses the importance of visual learning for future vision systems, which must automatically learn object models without human assistance. Biological systems use appearance for recognition, while machine vision systems lack learning capabilities. Recent research includes methods like regularization networks, Hebbian relaxation, and principal component analysis for face recognition. The paper proposes a technique for learning 3-D objects from 2D images by using eigenspace representation. The eigenspace is optimal for correlation and allows efficient recognition. The method involves generating a large set of images, compressing them into eigenvectors, and using the eigenspace for recognition. The paper concludes with a discussion of the proposed methodology's implications for future vision systems.This paper presents a method for automatically learning 3-D object models from their appearance in 2D images. The appearance of an object is influenced by its shape, reflectance, pose, and illumination. Recognition is based on matching appearance rather than shape. A compact representation of object appearance is proposed, parametrized by pose and illumination. For each object, a large set of images is generated by varying pose and illumination. These images are compressed into a low-dimensional subspace, called the eigenspace, where the object is represented as a manifold. Given an unknown image, the system projects it into the eigenspace and recognizes the object based on its position on the manifold. Experiments show that with less than 20 eigenspace dimensions, recognition is accurate with an average pose error of about 1 degree. A near real-time recognition system with 20 complex objects has been developed. The paper discusses the importance of visual learning for future vision systems, which must automatically learn object models without human assistance. Biological systems use appearance for recognition, while machine vision systems lack learning capabilities. Recent research includes methods like regularization networks, Hebbian relaxation, and principal component analysis for face recognition. The paper proposes a technique for learning 3-D objects from 2D images by using eigenspace representation. The eigenspace is optimal for correlation and allows efficient recognition. The method involves generating a large set of images, compressing them into eigenvectors, and using the eigenspace for recognition. The paper concludes with a discussion of the proposed methodology's implications for future vision systems.