Methods and systems for large-scale determination of RGBD camera poses
A method for refining poses includes receiving a plurality of poses and performing volumetric reconstruction for the plurality of poses to produce a global geometry. The method also includes refining each pose of the plurality of poses with respect to the global geometry to produce a plurality of refined poses.
1. A method for refining poses, the method comprising:
(a) receiving a plurality of poses each associated with one of a plurality of captured depth images, wherein each of the plurality of captured depth images is associated with one of a plurality of captured depth maps;
(b) performing volumetric reconstruction for the plurality of poses to produce a polygonal mesh corresponding to a global geometry;
(c) rendering, using the polygonal mesh, a depth map associated with each of the plurality of poses;
(d) aligning each captured depth map of the plurality of captured depth maps to each rendered depth map associated with each of the plurality of poses to provide refined poses; and
(e) iterating on (b), (c), and (d) using the refined poses for a predetermined number of iterations.
2. The method of claim 1 , wherein performing volumetric reconstruction comprises:
performing volumetric fusion; and
extracting the polygonal mesh.
3. The method of claim 2 , wherein performing volumetric reconstruction further comprises performing volumetric hole filling.
4. The method of claim 2 , wherein performing volumetric fusion comprises fusing each captured depth map associated with each of the plurality of poses into a truncated signed distance function.
5. The method of claim 4 , wherein extracting the polygonal mesh comprises converting the truncated signed distance function into a triangle mesh.
6. The method of claim 2 , wherein extracting the polygonal mesh comprises applying marching cubes.
7. The method of claim 1 , wherein aligning each rendered depth map associated with each of the plurality of poses to each captured depth map comprises an iterative closest point based alignment.
8. The method of claim 1 , wherein aligning each rendered depth map to each captured depth map comprises converging each rendered depth map to each captured depth map.
9. The method of claim 1 , wherein performing volumetric reconstruction for the plurality of poses and refining each pose of the plurality of poses with respect to the global geometry is iteratively performed based on a threshold for differences in one or more refined poses of plurality of refined poses and a corresponding pose of the plurality of poses.