3D Gaussian splatting for shaky vacation videos

What it does LongSplat turns unposed, unstructured long videos into explorable 3D Gaussian splat scenes. It jointly estimates camera poses and builds the scene incrementally, targeting the messy reality of consumer footage rather than controlled captures.

The interesting bit The system uses MASt3R for pose estimation and an adaptive octree anchoring scheme to keep memory sane across hundreds or thousands of frames. There’s also a conversion path to standard 3DGS format, though the README warns this trades quality and model size for viewer compatibility.

Key highlights

• Handles “casual” long videos without pre-computed camera poses
• Incremental joint optimization of poses and Gaussians
• Adaptive octree anchoring for memory efficiency at scale
• Converter to standard 3DGS point_cloud.ply for general viewers
• Released datasets: Free, Hike, and Tanks and Temples (preprocessed)
• ICCV 2025; NVIDIA-affiliated research

Caveats

• The README notes dynamic objects and insufficient frame overlap break things
• Conversion to standard 3DGS format degrades quality; pre-pruning recommended
• A December 2025 update fixed “training and eval pipeline for reproduction results” — suggesting earlier releases had reproducibility rough edges

Verdict Worth a look if you’re building 3D reconstruction pipelines for real-world video, or researching scalable Gaussian splatting. Skip if you need production-grade handling of dynamic scenes or want drop-in quality without tuning post_iter by hand.