Face Recognition from Video by Matching Image Sets

Abstract

As opposed to still-image based paradigms, video-based face recognition involves identifying a person from a video input. In video-based approaches, face detection and tracking are performed together with recognition, as usually the background is included in the video and the person could be moving or being captured unknowingly. By detecting and raster-scanning a face sub-image to be a vector, we can concatenate all extracted vectors to form an image set, thus allowing the application of face recognition algorithms based on matching image sets. It has been reported that linear subspace-based methods for face recognition using image sets achieve good recognition results. The challenge that remains is to update the linear subspace representation and perform recognition on-the-fly so that the recognitionfrom-video objective is not defeated. Here, we demonstrate how this can be achieved by using a well-studied incremental SVD updating procedure. We then present our online face recognition-from-video framework and the recognition results obtained.

Probabilistic Matching of Image Sets for Video-Based Face Recognition

“…Yamaguchi et al [7] propose linear subspaces as the representations and measure the similarities using canonical angles between subspaces. Following this work, Tat-Jun et al [4] apply an incremental SVD to compute linear subspaces on-line and use chordal distances as the distance metrics. Subspace-tosubspace distances have also been analyzed in the framework of Grassmannian manifolds [8] i.e.…”

“…With many observations from such frames, improvements might be obtained in the recognition stage. Techniques such as score fusion [3] and image set matching [4] have proven to be useful in disambiguating the decision choices. The availability of multiple observations may also mitigate the effects of non-optimal viewing conditions as well as inaccurate localization and feature extraction.…”