VRNet: Learning the Rectified Virtual Corresponding Points for 3D Point Cloud Registration

3D point cloud registration is fragile to outliers, which are labeled as the points without corresponding points. To handle this problem, a widely adopted strategy estimate relative pose based only on some accurate correspondences, achieved by building correspondences identified inliers or selecting reliable ones. However, these approaches usually complicated and time-consuming. By contrast, virtual point-based methods learn (VCPs) for all source uniformly distinguishing outliers inliers. Although time-efficient, learned VCPs exhibit serious collapse degeneration due insufficient supervision inherent distribution limitation. In paper, we propose exploit best of both worlds present novel robust framework. We follow idea but new type called rectified (RCPs), defined set with same shape target. Hence, pair consistent clouds, i.e. RCPs, formed rectifying RCPs (VRNet), through between can be accurately obtained. Since target, input clouds registered naturally. Specifically, first construct initial using an estimated soft matching matrix perform weighted average target Then, design correction-walk module offset rectify effectively breaks limitation VCPs. Finally, develop hybrid loss function enforce geometry structure consistency ...