**5. Conclusions**

In this paper, we presented a deep heatmap regression approach for facial landmark detection. We employed FC-DenseNets to extract dense feature maps along with an explicit kernel convolution for early-stage facial shape prediction. Starting with a suitable shape in the first stage, the detected shapes were refined to match the ground-truth shape during the last stage of the architecture. Our local appearance initialization subnet pursued a heatmap regression approach convolved with kernel convolution to serve as a local detector of facial landmarks in the first stage and the dilated skip convolution subnet was carefully designed to increase the performance of our dense prediction architecture and accurate spatial information by aggregating multi-scale contextual information for the sake of refining the local prediction of the first subnet. The proposed method achieved superior, or at least comparable, performance in comparison to state-of-the-art methods for challenging datasets, including LFPW, HELEN, 300W and ALFW2000-3D.

**Author Contributions:** The work presented here was completed with collaboration among all authors. Conceptualization, S.C.; Methodology, S.C., J.-G.L.; software, S.C.; Validation, S.C., J.-G.L. and H.-H.P.; Formal analysis, J.-G.L.; Writing—original draft preparation, S.C.; Writing—review and editing, S.C., H.-H.P.; Visualization, S.C.; Supervision, H.H.P.; Funding acquisition, H.-H.P.

**Funding:** This research was supported by the Chung-Ang University Young Scientist Scholarship (CAYSS), the Ministry of Education (Project number: NRF-2016R1D1A1B03933895, Project name: Face recognition and searching robust in pose, illumination and expression utilizing video big data) and the Ministry of Trade, Industry and Energy (Project number: P0002397, Project name: Advanced Expert Training Program for Industrial Convergence of Wearable Smart Devices.

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