LongSplat: Robust Unposed 3D Gaussian Splatting for Casual Long Videos

LongSplat is a framework for novel view synthesis from casually captured long videos without known camera poses.

It jointly optimizes camera poses and 3D Gaussian Splatting (3DGS) to address pose drift, inaccurate geometry initialization, and memory limitations.

Key components include incremental joint optimization, a pose estimation module using learned 3D priors, and an Octree Anchor Formation mechanism that adaptively converts dense point clouds into anchors based on spatial density.

Experiments on Tanks and Temples, Free, and Hike datasets show state-of-the-art results, with improved rendering quality, pose accuracy, and efficiency.

For example, on the Free dataset, LongSplat achieves an average PSNR of 27.88 dB, SSIM of 0.85, and LPIPS of 0.17, outperforming baselines like CF-3DGS, which often fails due to out-of-memory issues.

On the Hike dataset, it achieves 25.39 dB PSNR, surpassing LocalRF (23.56 dB).

The method also reduces model size to about 101 MB and trains in about 1 hour on an NVIDIA RTX 4090, achieving 281.71 FPS.

Ablations confirm the importance of each component, and the method is limited to static scenes with fixed intrinsics.