Abstract

An increasing number of computer vision and pattern recog- nition problems require structured regression techniques. Problems like human pose estimation, unsegmented action recognition, emotion pre- diction and facial landmark detection have temporal or spatial output dependencies that regular regression techniques do not capture. In this paper we present continuous conditional neural fields (CCNF) – a novel structured regression model that can learn non-linear input-output de- pendencies, and model temporal and spatial output relationships of vary- ing length sequences. We propose two instances of our CCNF framework: Chain-CCNF for time series modelling, and Grid-CCNF for spatial rela- tionship modelling. We evaluate our model on five public datasets span- ning three different regression problems: facial landmark detection in the wild, emotion prediction in music and facial action unit recognition. Our CCNF model demonstrates state-of-the-art performance on all of the datasets used.