The paper addresses the problem of automatically learning object models for recognition and pose estimation, focusing on matching appearance rather than shape. The appearance of an object in a two-dimensional image is influenced by its shape, reflectance properties, pose, and illumination. The authors propose a compact representation of object appearance parametrized by pose and illumination. They generate a large set of images by varying pose and illumination and compress this set into a low-dimensional subspace, called the eigenspace, where the object is represented as a manifold. For an unknown input image, the system projects it into the eigenspace and recognizes the object based on its position on the manifold, which also determines the object's pose in the image. Experiments using complex objects show that accurate recognition and pose estimation can be achieved with less than 20 dimensions in the eigenspace, achieving an average pose estimation error of about 1.0 degree. The paper concludes with a discussion on the methodology and its potential applications.The paper addresses the problem of automatically learning object models for recognition and pose estimation, focusing on matching appearance rather than shape. The appearance of an object in a two-dimensional image is influenced by its shape, reflectance properties, pose, and illumination. The authors propose a compact representation of object appearance parametrized by pose and illumination. They generate a large set of images by varying pose and illumination and compress this set into a low-dimensional subspace, called the eigenspace, where the object is represented as a manifold. For an unknown input image, the system projects it into the eigenspace and recognizes the object based on its position on the manifold, which also determines the object's pose in the image. Experiments using complex objects show that accurate recognition and pose estimation can be achieved with less than 20 dimensions in the eigenspace, achieving an average pose estimation error of about 1.0 degree. The paper concludes with a discussion on the methodology and its potential applications.