Parametric temporal alignment for the detection of facial action temporal segments

In this paper we propose the very first weakly supervised approach for detecting facial action unit temporal segments. This is achieved by means of behaviour similarity matching, where no training of dedicated classifiers is needed and the input facial behaviour episode is compared to a template. The inferred temporal segment boundaries of the test sequence are those transferred from the template sequence. To this end, a parametric temporal alignment algorithm is proposed to align a single exemplar sequence to the test sequence. The proposed strategy can accommodate flexible time warp functions, does not need to exhaustively align all frames in both sequences, and the optimal warp parameters can be found by an efficient Gauss-Newton gradient descent search. We show that our approach produces the best results to date for the problem at hand, and provides a promising opportunity to studying facial actions from a new perspective.