Ordinal hyperplanes ranker with cost sensitivities for age estimation

Age estimation from face images: Human vs. machine performance

“…Better results than ours have been reported on FG-NET in [2], [3], and [28] with MAE of 4.5, 4.5, and 4.1 respectively. The best known performance on MORPH Album2 is 4.18 years MAE reported by Guo and Mu [12], which is slightly better than our MAE of 4.2.…”

A deep analysis on age estimation

“…MAE computes the average age deviation error in absolute terms, MAE = M i=1 |â i − a i |/M, witĥ a i the estimated age of the i-th sample, a i its real age and M the total of samples. CS is defined as the percentage of images for which the error e is no higher than a given number of years l, as CS (l) = M e≤l /M (Chang et al (2011);Weng et al (2013); Han et al (2013)). Related publications typically supply either an eleven-point curve for age deviations [0 − 10], or simply the value CS (5).…”

“…With regards to the learning algorithm, several approaches have been proposed, including, among others, Support Vector Machines / Regressors (Guo et al (2009);Han et al (2013);Chang et al (2011);Weng et al (2013)), neural networks (Lanitis et al (2004)) and their variant of Conditional Probability Neural Network (Geng et al (2013)), Random Forests (Montillo and Ling (2009)), and projection techniques such as Partial Least Squares (PLS) and Canonical Correlation Analysis (CCA), along with their regularized and kernelized versions (Guo and Mu (2011). An extensive comparison of these classification schemes for age estimation has been reported ; Huerta et al (2014)), and the advantageousness of CCA was demonstrated over others, both regarding accuracy and efficiency.…”

“…In age estimation from images, typically the first phase after pre-processing is to extract visual features which need to be (1) discriminative among different classes, (2) robust within the same class, and (3) with a minimal dimensionality. One class of methods relies on flexible shape and appearance models such as ASM (Active Shape Model) and AAM (Active Appearance Model) to model aging patterns (Chang et al (2011);Geng et al (2013Geng et al ( , 2007; Lanitis et al (2004)). Such statistical models capture the main modes of variation in shape and intensity observed in a set of faces, and allow to encode face signatures based on such characterizations.…”

“…The most typical ones consist of either hand-tuned multi-level filter banks (Guo et al (2009) ;Geng et al (2013); Han et al (2013)), that intend to emulate the behavior of primary visual cortex cells, or fine-grained facial meshes to accomplish precise alignment through dozens of facial landmarks (Chang et al (2011);Geng et al (2007); Lanitis et al (2004)). In any case, the resulting extraction schemes are difficult to replicate, and the high-dimensional visual descriptors in many cases take considerable time to be computed.…”