Image Manifold Learning
Image manifolds offer a perceptually meaningful model to organize images for certain types of natural image sets. However, unsupervised methods are limited in the situations where a discriminant factor needs to be recovered from an image set with multiple latent factors of variation. Meanwhile, supervised manifold learning approaches incorporate image labels that provide additional constraints to the relationships between images and can be robust against irrelevant factors. During the course of my PhD study, I was very interested in learning on image manifolds with weak supervision, which needs much less manual labeling effort than supervised methods. My work considered three variants of weakly supervised learning on image manifolds, when image labels do not explain all the latent factors of image variation, or are only partly available or highly corrupted.