Skip to main content
Back to timeline
arXivSource publication:

Triangle primitives go real-time: Facet-SLAM hits sensor rate on a Jetson and cuts mesh distance to 0.86 cm

Synopsis

Facet-SLAM brings triangle primitives into real-time RGB-D simultaneous localization and mapping, so the map itself triangulates directly into a connected mesh. The earlier triangle-based TS-SLAM relied on photometric tracking at only a few frames per second, too slow for on-device localization. The system fixes each triangle's centroid from fused depth and registers frames by GICP while restricting the mapper to non-translating updates; it runs at about 100 FPS on a desktop GPU and at sensor rate on a Jetson AGX Orin, and on Replica it lowers the Delaunay mesh chamfer distance from 1.55 to 0.86 cm.

Source-provided article image: Facet-SLAM: Real-Time Triangle Splatting SLAM
arXiv

Interpretation

Triangle-primitive maps can now be built in real time from an RGB-D camera, and the map itself triangulates directly into a connected mesh without a separate extraction or fusion stage. The earlier triangle-primitive SLAM method, TS-SLAM, followed early 3DGS SLAM in photometric tracking, which is accurate but runs at only a few frames per second and cannot localize on device. On three TUM RGB-D scenes the system averages 1.76 ATE against 1.57 for TS-SLAM, while reaching 113 FPS on a desktop GPU versus 1.5 FPS for TS-SLAM and 36 FPS on a Jetson AGX Orin.

Tracking and mapping are split into two duties: triangle positions come from depth observations, while appearance, shape and orientation are left to the mapper. Point-to-plane registration on the triangles' own normals is unreliable on real depth, because the mapper's composite loss pushes triangles away from the observed surface. In ablation, point-to-point registration gives 4.50 ATE on TUM RGB-D and point-to-plane 8.49, against 1.76 for GICP with covariances from neighboring centroid scatter; restricting the mapper to non-translating updates gives 1.76 ATE on TUM RGB-D versus 5.11 when vertices move freely.

Map geometry beats both earlier triangle-splatting and 3DGS SLAM methods while rendering quality stays on par with triangle-splatting methods. 3DGS maps have poor surface geometry, and a good mesh needs post-hoc extraction, an intermediate representation or volumetric fusion. On Replica the Delaunay mesh chamfer distance is 0.86 cm against 1.55 cm for TS-SLAM, and the TSDF-fusion chamfer is 0.67 cm against 0.95 cm.

Perspective

The demonstrated scope is tracking, mapping and mesh quality with an RGB-D camera on TUM RGB-D, Replica and ScanNet, on an RTX 4090 desktop and a Jetson AGX Orin; the real-world experiment is a handheld camera moved around a desk structure, mapped entirely online on the Jetson AGX Orin.

Mapping quality overall remains below 3DGS methods, partly from design choices for computational efficiency such as 0th-degree spherical harmonics and more conservative densification; tracking accuracy is marginally below TS-SLAM, with average ATE of 1.76 versus 1.57 on TUM RGB-D.