This paper presents a new approach for face recognition based on the fusion of tensors of census transform histograms from Local Gaussian features maps. Local Gaussian feature maps encode the most relevant information from Gaussian derivative features. Census Transform (CT) histograms are calculated and concatenated to form a tensor for each class of Gaussian map. Multi-linear Principal Component Analysis (MPCA) is applied to each tensor to reduce the number of dimensions as well as the correlation between neighboring pixels due to the Census Transform. We then train Kernel Discriminative Common Vectors (KDCV) to generate a discriminative vector using the results of the MPCA. Results of recognition using MPCA of tensors-CT histograms from Gaussian features maps with KDCV is shown to compare favorably with competing techniques that use more complex features maps like for example Gabor features maps in the FERET and Yale datasets. Additional experiments were done in the Yale B+ extended Yale B Faces dataset to show the performance of Gaussian features map with hard illumination changes.

We present a new strategy for RANSAC sampling named BetaSAC, in reference to the beta distribution. Our proposed sampler builds a hypothesis set incrementally, selecting data points conditional on the previous data selected for the set. Such a sampling is shown to provide more suitable samples in terms of inlier ratio but also of consistency and potential to lead to an accurate parameters estimation. The algorithm is presented as a general framework, easily implemented and able to exploit any kind of prior information on the potential of a sample. As with PROSAC, BetaSAC converges towards RANSAC in the worst case. The benefits of the method are demonstrated on the homography estimation problem.

The half-octave Gaussian pyramid is an important tool in computer vision and image processing. The existence of a fast algorithm with linear computational complexity makes it feasible to implement the half-octave Gaussian pyramid in embedded computing systems using only integer arithmetic. However, the use of repeated convolutions using integer coefficients imposes limits on the minimum number of bits that must be used for representing image data. Failure to respect this limits results in serious degradation of the signal to noise ratio of pyramid images.In this paper we present a theoretical analysis of the accumulated error due to repeated integer coefficient convolutions with the binomial kernel. We show that the error can be seen as a random variable and we deduce a probabilistic model that describes it. Experimental and theoretical results demonstrate that the linear complexity algorithm using integer coefficients can be made suitable for video rate computation of a half-octave pyramid on embedded image acquisition devices.