Facial landmark point localization is a typical problem in computer vision and is extensively used for increasing accuracy of face recognition, facial expression analysis, face animation etc. In recent years, substantial effort have been deployed by many researcher to design a robust facial landmark detection system. However, it still remains as one of the most challenging tasks due to the existence of extreme poses, exaggerated facial expression, unconstrained illumination, etc. In this paper, we propose a novel coarse-to-fine deep recurrent-neural-network (RNN) based framework, which uses heat-map images for facial landmark point localization. The use of heat-map images allows us using the entire face image instead of the face initialization bounding boxes or patch images around the landmark points. Performance of our proposed framework shows significant improvement in case of handling difficult face images with higher degree of occlusion, variation of pose, large yaw angles and illumination. In comparison with the best current state-of-the-art technique a reduction of 45% in failure rate and an improvement of 11.5% in area under the curve for 300-W private test set are some of the main contributions of our proposed framework.