Datasets
Code
The LEAR programme detects 20 fiducial facial points. Instead of scanning an image or image region for the location of a facial point, it can use every location in a point's neighbourhood to predict where the target point is relative to that location. This considerably speeds up point detection. In order to obtain a precise prediction of the target location, a stochastic process decides the local patches used to sample from. Different target predictions are combined into a function by aggregation of Gaussian distributions, and the target corresponds to the maximum of this function. In this way, the effect of incorrect estimates is cancelled out, as correct estimates agree on the prediction while incorrect estimates produce more uncorrelated predictions. In only a few iterations the facial point can be found. Markov networks are used to confirm whether a set of predicted facial points adhere to shape statistics, and if they don't the Markov nets suggest new locations to continue the search for the facial point.
More information about the point detector can be found in:
B. Martinez, M. F. Valstar, X. Binefa, M. Pantic. Local Evidence Aggregation for Regression Based Facial Point Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence. 35(5): pp. 1149 - 1163, 2013.
We kindly request you to cite this work if you decide to use the point detector for research purposes.
Please note that we have re-annotated the BioID images as our point criterion, for which we already had abundant annotations, was different for some of the points.
The detector can be downloaded from the following links:
The BoRMaN programme detects 20 fiducial facial points. Instead of scanning an image or image region for the location of a facial point, it can use every location in a point's neighbourhood to predict where the target point is relative to that location. This considerable speeds up point detection. In only a few iterations the facial point can be found. Markov networks are used to confirm whether a set of predicted facial points adhere to shape statistics, and if they don't the Markov nets suggest new locations to continue the search for the facial point.
More information about the point detector can be found in:
Michel F. Valstar, Brais Martinez, Xavier Binefa, and Maja Pantic 'Facial Point Detection using Boosted Regression and Graph Models', IEEE Int'l Conf. Computer Vision and Pattern Recognition, pp. 2729-2736, San Francisco, USA, June 2010
We kindly request you to cite this work if you decide to use the point detector for research purposes.
The detector can be downloaded from the following links: