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

Recent work has made significant progress on using implicit functions, as a\ncontinuous representation for 3D rigid object shape reconstruction. However,\nmuch less effort has been devoted to modeling general articulated objects.\nCompared to rigid objects, articulated objects have higher degrees of freedom,\nwhich makes it hard to generalize to unseen shapes. To deal with the large\nshape variance, we introduce Articulated Signed Distance Functions (A-SDF) to\nrepresent articulated shapes with a disentangled latent space, where we have\nseparate codes for encoding shape and articulation. We assume no prior\nknowledge on part geometry, articulation status, joint type, joint axis, and\njoint location. With this disentangled continuous representation, we\ndemonstrate that we can control the articulation input and animate unseen\ninstances with unseen joint angles. Furthermore, we propose a Test-Time\nAdaptation inference algorithm to adjust our model during inference. We\ndemonstrate our model generalize well to out-of-distribution and unseen data,\ne.g., partial point clouds and real-world depth images.\n