Automatic localization of facial landmarks from expressive images of high complexity
Yulia Gizatdinova, Veikko Surakka · Tampere University Institutional Repository (Tampere University) · 2008
The aim of this study was to develop a fully automatic feature-based method for expression-invariant localization of facial landmarks from static facial images.It was a continuation of our earlier work in which we found that lower face expressions deteriorated the feature-based localization of facial landmarks the most.Taking into account the found crucial facial behaviours, the method was improved so that it allowed facial landmarks to be fully automatically located from expressive images of high complexity.Information on local oriented edges was utilized to compose edge maps of the image at several levels of resolution.Landmark candidates which resulted from this step were further verified by matching them against the orientation model.The present novelty came from two last steps of the method which were edge projection that was used to enhance a search for landmark candidates, and structural correction of the found candidates based on a face geometry model.The method was tested on three facial expression databases which represented a wide range of facial appearances in terms of prototypical facial displays and individual facial muscle contractions.To evaluate the localization results, a new performance evaluation measure was introduced that represented a localization result as a rectangular box instead of a conventional point representation.The evaluation of the method by the proposed rectangular and conventional point evaluation measures demonstrated a high overall performance of the method in localization of facial landmarks from images showing complex expressions in upper and lower face.