by M Wimmer, C Mayer and B Radig
Abstract:
Efficient and accurate localization of the components of human faces, such as skin, lips, eyes, and brows, provides benefit to various real-world applications. However, high intra-class and small inter-class variations in color prevent simple but quick pixel classifiers from yielding robust results. In contrast, more elaborate classifiers consider shape or region features but they do not achieve real-time performance. In this paper, we show that it definitely is possible to robustly determine the facial components and achieve far more than real-time performance. We also use quick pixel-level classifiers and provide them with a set of pixel features that are adapted to the image characteristics beforehand. We do not manually select the pixel features and specify the calculation rules. In contrast, our idea is to provide a multitude of features and let the Machine Learning algorithm decide which of them are important. The evaluation draws a comparison to fixed approaches that do not adapt the computation of the features to the image content in any way. The obtained accuracy is precise enough to be used for real-world applications such as for model-based interpretation of human faces.
Reference:
Robustly Classifying Facial Components Using a Set of Adjusted Pixel Features (M Wimmer, C Mayer and B Radig), In Proc. of the International Conference on Face and Gesture Recognition (FGR08), 2008.
Bibtex Entry:
@inproceedings{wimmer_robustly_2008,
author = {M Wimmer and C Mayer and B Radig},
title = {Robustly Classifying Facial Components Using a Set of Adjusted Pixel
Features},
booktitle = {Proc. of the International Conference on Face and Gesture Recognition
({FGR08)}},
year = {2008},
address = {Amsterdam, Netherlands},
month = {sep},
abstract = {Efficient and accurate localization of the components of human faces,
such as skin, lips, eyes, and brows, provides benefit to various
real-world applications. However, high intra-class and small inter-class
variations in color prevent simple but quick pixel classifiers from
yielding robust results. In contrast, more elaborate classifiers
consider shape or region features but they do not achieve real-time
performance. In this paper, we show that it definitely is possible
to robustly determine the facial components and achieve far more
than real-time performance. We also use quick pixel-level classifiers
and provide them with a set of pixel features that are adapted to
the image characteristics beforehand. We do not manually select the
pixel features and specify the calculation rules. In contrast, our
idea is to provide a multitude of features and let the Machine Learning
algorithm decide which of them are important. The evaluation draws
a comparison to fixed approaches that do not adapt the computation
of the features to the image content in any way. The obtained accuracy
is precise enough to be used for real-world applications such as
for model-based interpretation of human faces.},
keywords = {facial expressions},
}