vivit

Definition

Video Vision Transformer: a ViT adapted to video that embeds tokens as tubelets — small patches extended across consecutive frames — rather than as per-frame patches. The paper systematically compares factorised designs, including separate spatial and temporal encoders and factorised self-attention, trading accuracy against cost. It belongs to the same generation as TimeSformer, and both share the strategy of splitting attention across the spatial and temporal axes.