Towards automated selection of data fusion techniques

An investigation of multi-modal fusion schemes is done using synthetic data generation to determine how the data characteristics influence fusion. The goal is to select the best fusion scheme using data characteristics. Preliminary results are presented here that compare data concatenation to Kernel fusion in the presence of increasing dimensionality, linear/nonlinear decision boundaries and correlations between different modality features. It is found that data concatenation is better than Kernel fusion in low dimensions in general. It is also found that Kernel fusion is better than data concatenation when the optimal decision boundary is non-linear, and the dimensions are high. Correlations between modalities determine the information content, and Kernel fusion reduces the information content most when there is negative correlation between modalities. These results are applied to fingerprint live-ness detection on the ATVS database having three sensor modalities. As there are few features used per modality and the overall dimensionality is low, it is expected and confirmed that data concatenation is better than Kernel fusion.