A semi-supervised video transformer is a machine learning framework that applies transformer-based neural network architectures to video analysis tasks while training on a combination of limited labeled video data and a larger pool of unlabeled video data. By utilizing self-attention mechanisms across spatial and temporal dimensions, the model captures long-range dependencies and complex dynamics across video frames without relying entirely on costly manual annotations. These systems typically integrate semi-supervised learning strategies, such as pseudo-labeling, consistency regularization, and teacher-student training schemes, alongside video-specific temporal and spatial augmentations to generate reliable supervision signals from unlabeled clips. This approach reduces the data annotation bottleneck in video understanding domains, such as action recognition, while maintaining high predictive accuracy.
