Background and Question. One issue regarding the role of facial structure in visual processing concerns the relative importance of different face parts / features for recognition. However, most of the time research is confined to a restricted set of features, selected in advance and manually marked by the experimenters. Beyond the subjectivity of the approach, this often makes similar studies not comparable because of a different way of selecting and marking features. Also it makes the analysis of large datasets impractical. An automatic method for feature identification which parallels human performance both with respect to stimulus processing and with respect to processing outcome could solve the difficulties mentioned above and could lead to more principled approaches addressing the role of facial structure.

Methods. Automatic and manual image segmentation.

Results and Discussion. We designed a mutiple-cue patch-based segmentation algorithm and applied it to the 200 stimuli of the MPI dataset of color front-view Caucasian faces. The algorithm was trained with ground-truth data obtained from manual segmentations. We evaluated automatic segmentations and we found them to produce a reasonably good fit to human segmentations.

Current and future work. The ability of the method to deal with different sources of variations, e.g. viewpoint or pose, is critical for its utility. Future work needs to address robustness as well as flexibility issues, e.g. segmentation of the same stimulus at different levels of detail.

People. Adrian