Metric imitation by manifold transfer for efficient vision applications

Abstract

Metric learning has proved very successful. However, human annotations are necessary. In this paper, we propose an unsupervised method, dubbed Metric Imitation (MI), where metrics over cheap features (target features, TFs) are learned by imitating the standard metrics over more sophisticated, off-the-shelf features (source features, SFs) by transferring view-independent property manifold structures. In particular, MI consists of: 1) quantifying the properties of source metrics as manifold geometry, 2) transferring the manifold from source domain to target domain, and 3) learning a mapping of TFs so that the manifold is approximated as well as possible in the mapped feature domain. MI is useful in at least two scenarios where: 1) TFs are more efficient computationally and in terms of memory than SFs; and 2) SFs contain privileged information, but are not available during testing. For the former, MI is evaluated on image clustering, category-based image retrieval, and instance-based object retrieval, with three SFs and three TFs. For the latter, MI is tested on the task of example-based image super-resolution, where high-resolution patches are taken as SFs and low-resolution patches as TFs. Experiments show that MI is able to provide good metrics while avoiding expensive data labeling efforts and that it achieves state-of-the-art performance for image super-resolution. In addition, manifold transfer is an interesting direction of transfer learning.

Fast Algorithms for Linear and Kernel SVM+

“…We are also aware that there are some works proposed for solving the tasks in which training and test data contains different information, which are less related to SVM+ [2,3,5,7,13,14,15,20,30,35].…”

“…For example, the training images of many datasets for image recognition are annotated with privileged information such as attributes, object bounding boxes, textual descriptions, depth information. Although the raw test images in the real-world applications are not associated with such information, it has been demonstrated that such information is useful for learning classifiers with better recognition performance [5,11,20,29,32,33,35].…”

Transferring Knowledge Fragments for Learning Distance Metric from a Heterogeneous Domain

“…We conduct experiments on various popular applications: scene classification, object recognition, image retrieval and face verification. In addition to the Euclidean (EU) and single domain DML baselines, we further compare with several representative heterogeneous transfer learning approaches that could learn distance metric [5], [15], [17], [18]. The results validate the effectiveness of the proposed HTDML.…”

“…There exists a recent work of metric imitation [5], which learns an improved distance metric for the target features by utilizing them to approximate the manifold of source domain. It is more flexible than the other HTL approaches since the graph of the source domain can be computed offline, but the solution is obtained by performing eigenvalue decomposition on a large size graph matrix.…”

Face Hallucination by Learning Local Distance Metric

“…In [10], a metric was learnt to make the k-nearest neighbors belonging to the same class separated from examples in other classes by a large margin. In [11], a distance metric over cheap features was learned by the supervision of standard metric over more sophisticated source features.…”