Celebrity Face-Name Association in Web Videos using Unsupervised Approach

Computer Engineering

This paper explores the problem of missing name and missing faces in unconstrained videos with user provided metadata. Rather than depending upon supervised learning, a better relationship built from the content of a video, those relationship includes the arrival of faces in different spatio-temporal contexts and visual similarities between faces. The knowledge base consists of tagged images along with a set of names and celebrity social networks. Celebrity social network is built based on the co-occurrence statistics of celebrities in video metadata. Merging of relationship along with knowledge base is carried out via conditional random field. Two types of face-name association are investigated: within video face labeling and between video face labeling. The within video labeling takes care of noisy as well as incomplete labels in metadata, in which null assignment for the labels is permitted. Furthermore Between video face labeling addresses the flaws within metadata, particularly to correct incorrect names and label faces having no available names in metadata of a video. To do so it considers a gathering of socially associated videos for combined name inference. The experimental result analysis on web video dataset shows that proposed approach is very much powerful for handling the issue of missing names and incorrect names in face labeling problem than existing approaches.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.