keypointsimage.png
Student Project

Facial Keypoints Detection: An Effort to Top the Kaggle Leaderboard

Our team participated in the Facial Keypoints Recognition Kaggle competition.  The goal of this competition is to detect 15 different keypoints on the face, e.g. center of left eye, center of right eye, nose tip.  We built a convolutional neural network to tackle the task.  Before training the model, we used histogram stretching, Gaussian blurring, and image flipping to engineer the features of the training images.

The submissions to the competition are measured by the Root Mean Squared Error (RMSE).  That is, it measures the average distance of each team's model predictions from the true coordinates of the 15 keypoints on the face.  For example, an RMSE of 3.00 means that, on average, the predictions are three pixel away from the true keypoint.

Our best model measured an RMSE of 2.91784 on the Kaggle test data.  At the time of our submission, we earned a ranking of 13th on the Kaggle leaderboard.  Please use the following link to see the most recent standings.  Our team name is Lei&Marguerite&Younghak&Chris: 

Last updated: October 7, 2016