A research team spearheaded by Yongtao Wang from the Wang Xuan Institute of Computer Technology at Peking University, in partnership with VGI Labs and the University of California, Merced, has unveiled DynStream—a groundbreaking Gaussian feedforward reconstruction framework tailored for large-scale, pose-free streaming dynamic scenes. This innovative framework tackles three key challenges inherent in online 4D reconstruction: incremental processing, ensuring global consistency across different time frames, and achieving high-fidelity representation of dynamic scenes.

Leveraging three core modules—block-based local online reconstruction, global aggregation without the need for additional computation, and dynamic-static separation for efficient scene accumulation—DynStream enables feedforward output of camera trajectories. It continuously updates renderable dynamic scene representations by solely utilizing pose-free RGB online video stream input. This approach eliminates the necessity for reprocessing the entire scene or optimizing it on a scene-by-scene basis.

Experimental results demonstrate that, across multiple datasets encompassing both indoor and outdoor dynamic and static scenes, DynStream outperforms existing state-of-the-art (SOTA) feedforward Gaussian reconstruction methods in terms of rendering metrics. It effectively mitigates the issue of dynamic object ghosting. Even when applied to the longest 5km KITTI sequences, DynStream significantly reduces trajectory reconstruction errors, thereby achieving superior long-range scene and trajectory reconstruction quality.