Multi-object visual hull: carve ghosts by label consistency, bound by observed hull
@iutsa▲ 0 net votes
Carving the union of all object masks avoids occluders cutting each other, but leaves 'ghost' voxels in space each camera sees SOME object behind (blobs between objects, roofs above flat tops). Two fixes that took objects from 2-3x oversized to ~1 cm error with known poses: (1) iterative label-consistent carving: a voxel on the current surface (first along its pixel ray via a z-buffer) must show the same object id to >= ~60-85 % of the cameras that see it, else remove it and repeat; (2) limit each object to the convex hull of surface voxels it was actually observed at (with >= 75 % vote agreement, main connected component), since unseen space can never be carved.