The automatic analysis of emotion remains a challenging task in unconstrained experimental conditions. In this paper, we present our contribution to the 6th Audio/Visual Emotion Challenge (AVEC 2016), which aims at predicting the continuous emotional dimensions of arousal and valence. First, we propose to improve the performance of the multimodal prediction with low-level features by adding high-level geometry-based features, namely head pose and expression signature. The head pose is estimated by fitting a reference 3D mesh to the 2D facial landmarks. The expression signature is the projection of the facial landmarks in an unsupervised person-specific model. Second, we propose to fuse the unimodal predictions trained on each training subject before performing the multimodal fusion. The results show that our high-level features improve the performance of the multimodal prediction of arousal and that the subjects fusion works well in unimodal prediction but generalizes poorly in multimodal prediction, particularly on valence.