Face Recognition and 3D Face Modeling Technical Overview Overview Microsoft Research has developed a framework for face recognition. The system takes as input a simple frontal photograph of a face taken with normal illumination and neutral expression. It then creates a 3D face model. Different, realistic virtual faces with variant pose, illumination and expression are synthesized and used for face recognition. Test results showed that face recognition accuracy for non frontal poses is better than conventional approaches. The best performance improvements are achieved in half-profile views. In a first step frontal face detection and alignment algorithms are utilized to locate a frontal face and the facial feature points within an image, such as the contour points of the face, left and right eyes, mouth and nose. Then, the 3D face shape is reconstructed according to the feature points and a 3D face database. After that, the face model is texture-mapped by projecting the input 2D image onto the 3D face shape. Based on this 3D face model, virtual samples with variant pose, illumination and expression are synthesized to represent the 2D face image space. Finally, face recognition is conducted in this enlarged face subspace. The output images with variant pose, illumination and expression greatly improve face recognition results. The images shown below represent sample output of the 3D face modeling tool. The first and third rows are real photographs of various poses taken from a test database provided by Carnegie Mellon University. The second and fourth rows are the corresponding poses 3D Face Modeling generated. © 2006 Microsoft Corporation. All rights reserved. The information contained in this document relates to Microsoft Research pre-release/prototype products and technologies, which may be substantially modified before their final commercial release by third party IP venture companies. Accordingly, the information may not accurately describe or reflect the products and technologies when first commercially released. This document is provided for informational purposes only and Microsoft makes no warranties, express or implied, with respect to this document or the information contained in it. Reproduction and redistribution of this document requires the express permission of Microsoft Corporation. The images shown below illustrate how the system creates facial expressions. Again, the first row is the expressions taken from a test database provided by Carnegie Mellon University. The second row shows the 3D Face Modeling generated expressions corresponding to the first line. © 2006 Microsoft Corporation. All rights reserved. The information contained in this document relates to Microsoft Research pre-release/prototype products and technologies, which may be substantially modified before their final commercial release by third party IP venture companies. Accordingly, the information may not accurately describe or reflect the products and technologies when first commercially released. This document is provided for informational purposes only and Microsoft makes no warranties, express or implied, with respect to this document or the information contained in it. Reproduction and redistribution of this document requires the express permission of Microsoft Corporation. Components The Face Recognition and 3D Face Modeling framework consists of two parts: 1. 2D-to-3D face reconstruction: Starting from a frontal face image of a person with normal illumination and neutral expression the software finds 83 key feature points such as face contour points, eye centers and nose contour that characterize the face. In another step the key feature points are used to compute the 3D face model. In a final step the face texture is extracted from the input image and mapped to the 3D face model. 2. Face recognition: To increase the accuracy of the face recognition algorithm, many face images are created by rotating the 3D face model. Test results showed that face recognition accuracy for non frontal poses is better than conventional approaches. The best performance improvements are achieved in half-profile views. Core Scenarios 1. 2. 3. 4. 3D face reconstruction and 3D face modeling can be applied to video and online gaming (avatars, personalization). This system could be applied to the following surveillance scenarios: i. In casinos to detect people who have been barred from gambling. ii. In correctional facilities to screen visitors. iii. In airports to identify aircrews or to detect criminals. An intelligent photo sorting application could use face recognition technology to automatically tag and sort photos. A user would tag a photo with a particular person once and the application could automatically tag all other photos featuring that person. Other areas where face recognition is useful include i. Document control (photos in passports or national ID cards) ii. Computer security (user access verification) iii. Time and attendance (entry/exit verification) © 2006 Microsoft Corporation. All rights reserved. The information contained in this document relates to Microsoft Research pre-release/prototype products and technologies, which may be substantially modified before their final commercial release by third party IP venture companies. Accordingly, the information may not accurately describe or reflect the products and technologies when first commercially released. This document is provided for informational purposes only and Microsoft makes no warranties, express or implied, with respect to this document or the information contained in it. Reproduction and redistribution of this document requires the express permission of Microsoft Corporation. iv. Transactional authentication (credit cards, ATMs, point-of-sale) Documentation The following publication by Microsoft Research is related to Face Recognition and 3D Face Modeling: Yuxiao Hu, Dalong Jiang, Shuicheng Yan, Lei Zhang, Hongjiang Zhang, Automatic 3D Reconstruction for Face Recognition, in Proc. 6th IEEE International Conference on Automatic Face and Gesture Recognition (FG), Seoul, Korea, 2004 Related Work 1. 2. Face Recognition and 3D Face Modeling is based on face detection technology also available for licensing through the Microsoft IP Ventures program. Please click here for more information. Another application of the face alignment methods in Face Recognition and 3D Face Modeling is implemented in a technology called Personalized Facial Sketch also available for licensing. Please click here for more information. Differentiating Features 1. 2. 3. Only one single photo of a frontal face is required as input for Face Recognition. The burdensome work of adding faces to the database is significantly less compared to other face recognition solutions available in the market. The synthesized face samples provide the capability to conduct recognition under difficult conditions including variant pose, illumination and expression. The 2D-to-3D integrated face reconstruction approach is fully automatic and faster than other 3D reconstruction approaches. © 2006 Microsoft Corporation. All rights reserved. The information contained in this document relates to Microsoft Research pre-release/prototype products and technologies, which may be substantially modified before their final commercial release by third party IP venture companies. Accordingly, the information may not accurately describe or reflect the products and technologies when first commercially released. This document is provided for informational purposes only and Microsoft makes no warranties, express or implied, with respect to this document or the information contained in it. Reproduction and redistribution of this document requires the express permission of Microsoft Corporation. Demo/Prototype Development Status Face Recognition and 3D Face Modeling is currently a research prototype. The following screenshot shows the user interface for 3D Face Reconstruction and Synthesis: Technical Specifications Implemented in C++ Recognition speed: ~4 sec/face image (512 x 512 pixels) on a P4, 1.3 GHz PC with 256 MB RAM © 2006 Microsoft Corporation. All rights reserved. The information contained in this document relates to Microsoft Research pre-release/prototype products and technologies, which may be substantially modified before their final commercial release by third party IP venture companies. Accordingly, the information may not accurately describe or reflect the products and technologies when first commercially released. This document is provided for informational purposes only and Microsoft makes no warranties, express or implied, with respect to this document or the information contained in it. Reproduction and redistribution of this document requires the express permission of Microsoft Corporation.