Metric imitation by manifold transfer for efficient vision applications

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