Photo Tagging by Collection-Aware People Recognition

Some mature findings in person recognition research restrict their use cases for dealing with near frontal face poses under controlled environments conditions. Such special pose and environment conditions can be accepted if imposed as a requirement for security applications, but they are not reasonable constraints when dealing with personal collections. In this paper a person recognition approach is proposed designed for aiding the task of personal collections tagging under non-controlled conditions. It can be generally described as a collection-aware clustering approach, that evaluates people appearance across the photo set. A classificatory procedure working over a trained Kohonen network is proposed. The network is responsible for learning people appearance, working over a dynamically reduced feature space.