PKU, VGI Labs and UC Merced unveil DynStream for pose-free streaming 4D reconstruction
A team from Peking University's Wang Xuan Institute of Computer Technology, VGI Labs and UC Merced has proposed DynStream, an online streaming 4D Gaussian feed-forward reconstruction framework. It takes an RGB online video stream with no camera poses and feed-forward outputs a camera trajectory plus a continuously updated, renderable dynamic scene representation: video is processed in temporal chunks, local reconstructions are merged into a shared world frame using predicted relative poses and a per-chunk scale factor, and Gaussians are split into static and dynamic by a predicted dynamic probability. The authors report better rendering metrics than feed-forward baselines such as AnySplat, MVSplat and NeoVerse across five datasets including Waymo and SpatialVID, and a Sim(3) ATE reduced from hundreds of meters to 40 m on 11 KITTI sequences totaling 23,201 frames and up to 5 km long. These figures are author-reported and no independent replication is provided.
The work aims to replace per-scene optimization with pose-free feed-forward reconstruction for long online video, which if reproduced would affect how dynamic scene reconstruction and long-range camera trajectory estimation are done.
机器之心 2026-10-19 12:02 北京
从「无位姿长视频」到「可持续更新的动态三维场景」！
随手拍一段街景，镜头在移动，行人在穿行，街道上的车也在开。如果要边拍边重建这个变化的三维场景，麻烦就来了：未来画面还没到，如何分清哪些变化来自相机移动，哪些来自物体运动？随着镜头不断向前，如何将前一分钟重建的街道与后一分钟看到的路口准确拼接，避免场景错位或动态物体留下重影？当视频源源不断地输入时，系统又如何避免重复计算，让重建始终跟得上拍摄速度？
#DynStream#Peking University#VGI Labs#UC Merced#VDIG Lab#KITTI#Waymo#AnySplat#MVSplat#NeoVerse#SpatialVID