Chin-Yang Lin Cheng Sun Fu-En Yang Min-Hung Chen Yen-Yu Lin Yu-Lun Liu

Abstract

LongSplat addresses critical challenges in novel view synthesis (NVS) fromcasually captured long videos characterized by irregular camera motion, unknowncamera poses, and expansive scenes. Current methods often suffer from posedrift, inaccurate geometry initialization, and severe memory limitations. Toaddress these issues, we introduce LongSplat, a robust unposed 3D GaussianSplatting framework featuring: (1) Incremental Joint Optimization thatconcurrently optimizes camera poses and 3D Gaussians to avoid local minima andensure global consistency; (2) a robust Pose Estimation Module leveraginglearned 3D priors; and (3) an efficient Octree Anchor Formation mechanism thatconverts dense point clouds into anchors based on spatial density. Extensiveexperiments on challenging benchmarks demonstrate that LongSplat achievesstate-of-the-art results, substantially improving rendering quality, poseaccuracy, and computational efficiency compared to prior approaches. Projectpage: https://linjohnss.github.io/longsplat/