Triangle Splatting SLAM

Nicholas Fry1,2 Eric Dexheimer2 Kirill Mazur2 Paul H. J. Kelly1,2 Andrew J. Davison2
1Software Performance Optimisation Group, Imperial College London 2Department of Computing, Imperial College London
Paper Code Coming Soon Video BibTeX

A dense RGB-D SLAM system with differentiable triangles as the 3D map representation, for on-the-fly mesh extraction.

Scroll for more

Method

The system uses a differentiable triangle rasteriser for tracking and mapping. Triangles are spawned from back-projected depth maps, optimised photometrically, and can be converted to a connected mesh on demand. Unlike methods that use TSDF Fusion, we can track and edit the same unified triangle representation.

System diagram showing the tracking, keyframing, and triangle mapping stages connected via co-visibility checks, with a Delaunay output.

System overview. Tracking estimates the pose of each incoming RGB-D frame; keyframing selects high-co-visibility frames; mapping jointly refines poses and the triangle soup; Delaunay recovers a connected mesh on demand.

Demonstrations

Live captures from the system: photorealistic rendering, physics simulations, online mesh editing and tracking.

Citation

If you find this work useful, please consider citing:

@article{fry2026triangle, title = {Triangle Splatting {SLAM}}, author = {Fry, Nicholas and Dexheimer, Eric and Mazur, Kirill and Kelly, Paul H. J. and Davison, Andrew J.}, year = {2026}, eprint = {2605.31419}, archivePrefix = {arXiv}, primaryClass = {cs.CV} }

Acknowledgements

This research is funded by the EPSRC On-Sensor Computer Vision grant (EP/Y020499/1). Website inspired by KV-Tracker.