Early, intermediate and late fusion strategies for robust deep learning-based multimodal action recognition

Multimodal action recognition techniques combine several image modalities (RGB, Depth, Skeleton, and InfraRed) for a more robust recognition. According to the fusion level in the action recognition pipeline, we can distinguish three families of approaches: early fusion, where the raw modalities are combined ahead of feature extraction; intermediate fusion, the features, respective to each modality, are concatenated before classification; and late fusion, where the modality-wise classification results are combined. After reviewing the literature, we identified the principal defects of each category, which we try to address by first investigating more deeply the early-stage fusion that has been poorly explored in the literature. Second, intermediate fusion protocols operate on the feature map, irrespective of the particularity of human action, we propose a new scheme where we optimally combine modality-wise features. Third, as most of the late fusion solutions use handcrafted rules, prone to human bias, and far from real-world peculiarities, we adopt a neural learning strategy to extract significant features from data rather than assuming that artificial rules are correct. We validated our findings on two challenging datasets. Our obtained results were as good or better than their literature counterparts.