A method for part matching of two objects based on metric learning and universal domain description

In AI applications, one has to work not only with features, but also metric descriptions of objects. This requires the development of a special set of methods for constructing, converting, correcting and using metric descriptions. We provide a systematization of the methods. In particular, a new approach to the problem of part matching is considered. First, we propose to use the 'universal graph' as a way to enrich an individual matching problem with general information about the domain. This reduces the part matching problem to the graph matching problem. Second, we propose a fast graph matching method based on metric learning. At the same time, we generally avoid the quadratic assignment problem, which allows us to achieve high computational efficiency. Experiments demonstrate good performance compared to conventional methods.