This paper introduces StreamVGGT, a causal transformer model that reconstructs 4D spatial-temporal geometry from video in real-time. By caching historical tokens and using causal attention, it processes video frames incrementally, offering faster inference than traditional methods while maintaining competitive accuracy thanks to knowledge distillation from a more computationally expensive teacher model.

None identified

The paper presents a novel causal transformer architecture for streaming 4D visual geometry reconstruction, addressing the limitations of existing offline methods. The proposed StreamVGGT achieves competitive performance compared to state-of-the-art offline models while significantly reducing inference overhead, paving the way for real-time 4D vision systems. While some limitations regarding memory scalability and dependence on teacher model quality exist, the overall contribution and innovative approach warrant a strong rating.