arxiv Viewpoint Equivariance for Multi-View 3D Object Detection