Stéphane Pigeon

The M2VTS Multimodal Face Database (Release 1.00)

The primary goal of the M2VTS project is to address the issue of secured access to buildings or multi-media services by the use of automatic person verification based on multimodal strategies (secured access based on speech, face images and other information). This paper presents an overview of the multimodal face database recorded at UCL premises for the purpose of research applications inside the M2VTS project. This database offers synchronized video and speech data as well as image sequences allowing to access multiple views of a face. This material should permit the design and the testing of identification strategies based on speech andro labial analysis, frontal and/or profile face analysis as well as 3-D analysis thanks to the multiple views. The M2VTS Database is available to any non-commercial user on request to the European Language Resource Agency.

Image-based multimodal face authentication

Multimodality is a key word in order to increase the efficiency and the robustness of person authentication algorithms. Most of the multimodal authentication schemes currently developed, tend to combine speech and image-based features together and benefit from the high performance offered by the speech modality. Depending on the application, speech data is not always available or cannot be used. This paper takes these cases into account and investigates the best performance achievable through a system based on facial images only, using information extracted from both profile and frontal views. Starting from two different profile-related modalities, one based on the profile shape, the other on the grey level distribution along this shape, we will see how to issue a profile-based expert whose performance is improved compared to each profile modality taken separately. A second expert will use invariant parts of the frontal view in order to issue a frontal-based authentication. Different fusion schemes will be studied and the best approach will be applied in order to efficiently combine our two experts. This will result in a robust image-based person authentication scheme that offers a success rate of 96.5% on the M2VTS database

Profile Authentication Using a Chamfer Matching Algorithm

This paper investigates the use of the profile shape to recognize human faces. In order to test the intrinsic efficiency of profile authentication, we will make use of a method that directly works on the profile contour encoded as x-y coordinates. This way, the performance rated here will only depend on the ability of the profile view to dissociate faces and not on the choice of a given set of profile features and/or the quality of their extraction.

Multiple experts for robust face authentication

Most of the multimodal authentication schemes currently developed, combine speech and image-based features together and benefit from the high performance offered by the speech modality. Depending on the application, speech data is not always available or cannot be used. This paper takes these cases into account and investigates the best performance that can be achieved by a system based on facial images only, using information taken from both profile and frontal views. Starting from two different profile-related modalities, one based on the profile shape, the other on the grey level distribution along this shape, we will issue a first profile- based expert whose performance is improved compared to each profile modality taken separately. A second expert will use the most invariant part of the frontal view, namely information from a rectangular grey level window centered around the eyes and nose features, in order to issue a frontal-based authentication. Different fusion schemes are studied and the best approach will be applied in order to efficiently combine our two experts. This will result in a robust image-based person authentication scheme that offers a success rate of 96.5% measured on the M2VTS multimodal face database.