(This work was done in 2016.) This paper describes an approach to predicting keypoint positions on greyscale images of faces, as part of the Facial Detection (2016) Kaggle competition. Facial keypoints include centers and corners of the eyes, eyebrows, nose and mouth, among other facial features. Our methodology involves four steps to producing our output predictions. The first step involves deciding the criterion for model accuracy which is RMSE, the root means square error of all keypoint coordinates. For the second stage, we use principal components analysis to reduce the dimensionality of the image and edge-detection to better identify the key components needed to perform the facial keypoint detection. The we predict the missing feature through a neural network with a sigmoid non-linearity on the output layer. At training time, we will introduce a special convolutional neural network model which looks like an inverted hourglass and compare with the baseline model. We evaluate our different model architectures, with and without data augmentation techniques based on their Root Mean Squared Error scores they produce on the test set.
