**4. Conclusions**

In this paper, an ensemble of state-of-the-art face detectors is combined with a set of filters calculated from both the depth map and the color image. The filters reduce the number of false positives produced by the ensemble while maximizing the detection rate. A set of seven filters based on the size, the flatness, or the unevenness of the candidate face regions, or on the size of the larger cluster of the depth map of the candidate face regions, or on eye detection or the degree of ellipse fitting are evaluated, including a new method proposed here that is based on processing the candidate region with different wavelets. The method proposed in this work for developing an ensemble of face detectors uses the depth map to obtain increased effectiveness even under many indoor and outdoor illumination settings.

The experimental results demonstrate that the filtering steps significantly reduce the number of false positives (from 16,325 to 1018) without significantly decreasing the detection rate (from 92.02 to 90.07) on a challenging dataset containing images with cluttered and complicated backgrounds. The performance of the proposed system is also reported on the challenging BioID benchmark to validate the approach presented here further and to compare the best performing ensemble with the state-of-the-art in face detection.

The face detector named SFD is shown to outperform all other standalone methods. However, an ensemble proposed here that combines SFD with other types of face detectors is shown to boost the standalone performance of SFD. Obviously, increasing the number of face detectors included in ensembles increases the number of false positives; however, as the experiments in this work demonstrate, the application of a new cascade of filters reduces this number to acceptable levels.

**Author Contributions:** Conceptualization, L.N. and A.L.; methodology, L.N.; software, L.N. and A.L.; validation, L.N., S.B. and A.L.; formal analysis, L.N.; investigation, A.L.; resources, S.B.; writing—original draft preparation, A.L. and S.B.; writing—review and editing, S.B.; visualization, S.B. and A.L.

**Funding:** This research received no external funding.

**Conflicts of Interest:** The authors declare no conflict of interest.
