Truncated Signed Distance Function

Storing the signed distance to the nearest surface in a voxel grid, used to fuse multiple depth frames into a 3D model.

TSDF is a 3D scene representation: space is divided into small cubes (voxels), and each voxel stores its distance to the nearest object surface — positive in front of the surface, negative behind it — with values far from the surface truncated to a fixed limit, keeping only information near the surface itself. Curless and Levoy proposed fusing multiple depth maps with this kind of volumetric method in 1996, and KinectFusion made it real-time on a GPU in 2011. As each new depth frame arrives, its observations are weighted and averaged into the voxel grid according to the camera pose, and noise gets smoothed out as more frames accumulate; finally, the Marching Cubes algorithm extracts the zero-distance surface as a triangle mesh. TSDF is widely used for robot mapping, obstacle avoidance, and grasping, with implementations in Open3D and NVIDIA’s nvblox.

ExampleA handheld RGB-D camera is swept around a table; Open3D fuses each frame’s depth into a TSDF volume according to its pose, and finally exports a mesh model of the whole tabletop and the objects on it.

Also called

TSDF, TSDF Fusion, Truncated Signed Distance Field

Related

Signed Distance Field / Function · Voxel · Triangle Mesh · Euclidean Signed Distance Field · nvblox · Depth Map

Sources

Open3D: RGBD integration (TSDF volume)