DualPM: Dual Posed-Canonical Point Maps for 3D Shape and Pose Reconstruction

dualpm.github.io canonical point map 𝑸 posed point map 𝑷 𝛷 input image 𝑰 deformation field 𝑷 -𝑸 input image 𝑰 𝑷 -input view 𝑷 -novel views fitted skeleton Figure 1. Left: We map an image of an object to its Dual Point Maps (DualPMs), a pair of point maps P , defined in a camera space, and Q, defined in a canonical space where the object has a neutral pose. The pose is thus given by the flow P -Q. Right: The DualPMs are easy to predict with a neural network, enabling effective 3D object reconstruction and facilitating geometric tasks like detecting 3D keypoints and fitting a 3D skeleton. For visualization, we color each point with its coordinate in the canonical point maps.