Multi-modal Conditional Feature Enhancement for Facial Action Unit Recognition

Current state-of-the-art methods in multi-modal fusion typically rely on generating a new shared representation space onto which multi-modal features are mapped for the goal of obtaining performance improvements by combining the individual modalities. Often, these heavily fine-tuned feature representations would have strong feature discriminability in their own spaces which may not be present in the fused subspace owing to the compression of information arising from multiple sources. To address this, we propose a new approach to fusion by enhancing the individual feature spaces through information exchange between the modalities. Essentially, domain adaptation is learnt by building a shared representation used for mutually enhancing each domain’s knowledge. In particular, the learning objective is modeled to modify the features with the overarching goal of improving the combined system performance. We apply our fusion method to the task of facial action unit (AU) recognition by learning to enhance the thermal and visible feature representations. We compare our approach to other recent fusion schemes and demonstrate its effectiveness on the MMSE dataset by outperforming previous techniques.