Manifold representation of multi-view images

Images of the same object lie on a low-dimensional manifold (view manifold) in the visual space. View manifolds can be used to represent viewpoint variation of multi-view images in the embedding space, and can be very helpful to multi-view object detection, classification, and viewpoint estimation. In this paper, we introduce a conceptual manifold as a common representation of all view manifolds. In order to evaluate the performance of the conceptual manifold representation, we learn a generative model that can map from the manifold representation to visual inputs for the tasks of arbitrary view image synthesis and viewpoint estimation. We did experiments on COIL-20 dataset, and compared with popular manifold learning methods. Experimental results show that our conceptual manifold representation can effectively describe the viewpoint variation of multi-view images with strong robustness, and outperform the view manifolds learned by popular manifold learning methods.