HMSFU: A hierarchical multi-scale fusion unit for video prediction and beyond

Video prediction is the process of learning necessary information from historical frames to predict future video frames. Learning features from historical frames is a crucial step in this process. However, most current methods have a relatively single-scale learning approach, even if they learn features at different scales, they cannot fully integrate and utilise them, resulting in unsatisfactory prediction results. To address this issue, a hierarchical multi-scale fusion unit (HMSFU) is proposed. By using a hierarchical multi-scale architecture, each layer predicts future frames at different granularities using different convolutional scales. The abstract features from different layers can be fused, enabling the model not only to capture rich contextual information but also to expand the model's receptive field, enhance its expressive power, and improve its applicability to complex prediction scenarios. To fully utilise the expanded receptive field, HMSFU incorporates three fusion modules. The first module is the single-layer historical attention fusion module, which uses an attention mechanism to fuse the features from historical frames into the current frame at each layer. The second module is the single-layer spatiotemporal fusion module, which fuses complementary temporal and spatial features at each layer. The third module is the multi-layer spatiotemporal fusion module, which fuses spatiotemporal features from different layers. Additionally, the authors not only focus on the frame-level error using mean squared error loss, but also introduce the novel use of Kullback–Leibler (KL) divergence to consider inter-frame variations. Experimental results demonstrate that our proposed HMSFU model achieves the best performance on popular video prediction datasets, showcasing its remarkable competitiveness in the field.