Face Recognition with Kernel Class Dependent Feature Analysis using Correlation Filters (Face Recognition Grand Challenge)

Kernel class-dependent feature analysis (KCFA) method has been proposed to generalize the approaches using correlation filters for face recognition. The dimensionality of the test images (gallery and probe) is efficiently reduced by  projecting the test images  onto the class specific basis vectors designed  from the generic training set using correlation filters. We evaluate our proposed algorithm using the Face Recognition Grand Challenge Dataset showing better performance over other approaches such as PCA, LDA, KPCA and KLDA.

Visual and Thermal Face Recognition using Correlation Filters

Correlation filter designs have shown to be distortion invariant and the advantages of using thermal IR images are due to their invariance to visible illumination variations. A combined use of thermal IR image data and correlation filters makes a viable means for improving the performance of face recognition techniques, especially beyond visual spectrum.

Face Modeling from Structure From Motion

Structure from Motion (SfM) theory enables to estimate the 3D shape and the camera motions as well with images taken different views. Using m different frames and n correspondence points across frames, the Factorization Method estimates camera parameters and 3D shape in the scene with Singular Value Decomposition (SVD) method. The most challenging part in the SfM is so called the Correspondence Problem. Researchers have been trying to track those features using the Kalman Filter and the Lucas Kanade optical flow algorithm (LK tracker). Those tracking algorithms easily fail if the correspondent points not consistent across frames. Thus more reliable tracking algorithms are needed to be developed in order to recover 3D faces in a fully automated way.