虽然新颖的视图合成(NVS)在3D计算机视觉中取得了进步,但通常需要从密集的视点对摄像机内在和外部设备进行初始估计。这种预处理通常是通过结构 - 运动(SFM)管道来进行的,这是一种可以缓慢且不可靠的操作,尤其是在稀疏视图方案中,匹配的功能不足,无法进行准确的重建。In this work, we integrate the strengths of point-based representations (e.g., 3D Gaus- sian Splatting, 3D-GS) with end-to-end dense stereo mod- els (DUSt3R) to tackle the complex yet unresolved is- sues in NVS under unconstrained settings, which encom- passes pose-free and sparse view challenges.我们的框架工作,InstantsPlat,用3D-GS统一了密集的立体声先验,以构建稀疏场景的3D高斯大型场景 -