| Face detection | Go back to the index | Limitations |
As shown in the Face Detection section, several local maxima in the result image can correspond to a single face. This effect is mainly produced by the automatic color detection algorithm. The sum of the correlations produces a peak around the middle of the face. However, in that area many natural photos have the holes of the nose. Those holes do not have the skin color, and therefore they are attenuated by the automatic color detection algorithm. Therefore, when the result of the sigmoid is added to the sum of the combined correlations, instead of a single peak, multiple local maxima are obtained.
The automatic clustering algorithm detects all the clusters of points whose diameter is smaller than the head size (which is assumed to be around 40% bigger than the face size). The weighted average of the points of each cluster is calculated, resulting in the final face position estimations. The blue dot in the following image shows the weighted average of the local maxima, represented by the red dots. The green box is drawn around the final face position estimation, and its size is proportional to the estimated correct size.
| Face detection | Go back to the index | Limitations |