Convolutional Neural Nets (CNNs) have become the reference technology for many computer vision problems. Although CNNs for facial landmark detection are very robust, they still lack accuracy when processing images acquired in unrestricted conditions. In this paper we investigate the use of a cascade of Neural Net regressors to increase the accuracy of the estimated facial landmarks. To this end we append two encoder-decoder CNNs with the same architecture. The first net produces a set of heatmaps with a rough estimation of landmark locations. The second, trained with synthetically generated occlusions, refines the location of ambiguous and occluded landmarks. Finally, a densely connected layer with shared weights among all heatmaps, accurately regresses the landmark coordinates. The proposed approach achieves state-of-the-art results in 300W, COFW and WFLW that are widely considered the most challenging public data sets.