Margin-constrained multiple kernel learning based multi-modal fusion for affect recognition

Recent advances in multiple-kernel learning (MKL) show the effectiveness to fuse multiple base features in object detection and recognition. However, MKL tends to select only the most discriminative base features but ignore other less discriminative base features which may provide complementary information. Moreover, MKL usually employ Gaussian RBF kernels to transform each base feature to its high dimensional space. Generally, base features from different modalities require different kernel parameters for obtaining the optimal performance. Therefore, MKL may fail to utilize the maximum discriminative power of all base features from multiple modalities at the same time. In order to address these issues, we propose a margin-constrained multiple-kernel learning (MCMKL) method by extending MKL with margin constraints and applying dimensionally normalized RBF (DNRBF) kernels for application of multi-modal feature fusion. The proposed MCMKL method learns weights of different base features according to their discriminative power. Unlike the conventional MKL, MCMKL incorporates less discriminative base features by assigning smaller weights when constructing the optimal combined kernel, so that we can fully take the advantages of the complementary features from different modalities. We validate the proposed MCMKL method for affect recognition from face and body gesture modalities on the FABO dataset. Our extensive experiments demonstrate favorable results as compared to the existing work, and MKL-based approach.