Residual Attention-Based Fusion for Video Classification

Video data is inherently multimodal and sequential. Therefore, deep learning models need to aggregate all data modalities while capturing the most relevant spatio-temporal information from a given video. This paper presents a multimodal deep learning framework for video classification using a Residual Attention-based Fusion (RAF) method. Specifically, this framework extracts spatio-temporal features from each modality using residual attention-based bidirectional Long Short-Term Memory and fuses the information using a weighted Support Vector Machine to handle the imbalanced data. Experimental results on a natural disaster video dataset show that our approach improves upon the state-of-the-art by 5% and 8% regarding F1 and MAP metrics, respectively. Most remarkably, our proposed residual attention model reaches a 0.95 F1-score and 0.92 MAP for this dataset.