A-SDF: Learning Disentangled Signed Distance Functions for Articulated Shape Representation

  • 2021-04-15 17:53:54
  • Jiteng Mu, Weichao Qiu, Adam Kortylewski, Alan Yuille, Nuno Vasconcelos, Xiaolong Wang
  • 3

Abstract

Recent work has made significant progress on using implicit functions, as acontinuous representation for 3D rigid object shape reconstruction. However,much less effort has been devoted to modeling general articulated objects.Compared to rigid objects, articulated objects have higher degrees of freedom,which makes it hard to generalize to unseen shapes. To deal with the largeshape variance, we introduce Articulated Signed Distance Functions (A-SDF) torepresent articulated shapes with a disentangled latent space, where we haveseparate codes for encoding shape and articulation. We assume no priorknowledge on part geometry, articulation status, joint type, joint axis, andjoint location. With this disentangled continuous representation, wedemonstrate that we can control the articulation input and animate unseeninstances with unseen joint angles. Furthermore, we propose a Test-TimeAdaptation inference algorithm to adjust our model during inference. Wedemonstrate our model generalize well to out-of-distribution and unseen data,e.g., partial point clouds and real-world depth images.

 

Quick Read (beta)

loading the full paper ...