RepAttn3D: Re-parameterizing 3D attention with spatiotemporal augmentation for video understanding

The technique of structural re-parameterization has been widely adopted in Convolutional Neural Networks (CNNs) and Multi-Layer Perceptrons (MLPs) for image-related tasks. However, its integration with attention mechanisms in the video domain remains relatively unexplored. Moreover, video analysis tasks continue to face challenges due to high computational costs, particularly during inference. In this paper, we investigate the re-parameterization of widely-used 3D attention mechanism for video understanding by incorporating a spatiotemporal coherence prior. This approach allows the learning of more robust video features while introducing negligible computational overhead at inference time. Specifically, we propose a SpatioTemporally Augmented 3D Attention (STA-3DA) module as a building block for Transformer architectures. The STA-3DA integrates 3D, spatial, and temporal attention branches during training, serving as an effective replacement for standard 3D attention in existing Transformer models and leading to improved performance. During testing, the different branches are merged into a single 3D attention operation via learned fusion weights, resulting in minimal additional computational cost. Experimental results demonstrate that the proposed method achieves competitive video understanding performance on benchmark datasets such as Kinetics-400 and Something-Something V2.