eduzhai > Applied Sciences > Engineering >

HyperFlow Representing 3D Objects as Surfaces

  • Save

... pages left unread,continue reading

Document pages: 18 pages

Abstract: In this work, we present HyperFlow - a novel generative model that leverageshypernetworks to create continuous 3D object representations in a form oflightweight surfaces (meshes), directly out of point clouds. Efficient objectrepresentations are essential for many computer vision applications, includingrobotic manipulation and autonomous driving. However, creating thoserepresentations is often cumbersome, because it requires processing unorderedsets of point clouds. Therefore, it is either computationally expensive, due toadditional optimization constraints such as permutation invariance, or leads toquantization losses introduced by binning point clouds into discrete voxels.Inspired by mesh-based representations of objects used in computer graphics, wepostulate a fundamentally different approach and represent 3D objects as afamily of surfaces. To that end, we devise a generative model that uses ahypernetwork to return the weights of a Continuous Normalizing Flows (CNF)target network. The goal of this target network is to map points from aprobability distribution into a 3D mesh. To avoid numerical instability of theCNF on compact support distributions, we propose a new Spherical Log-Normalfunction which models density of 3D points around object surfaces mimickingnoise introduced by 3D capturing devices. As a result, we obtain continuousmesh-based object representations that yield better qualitative results thancompeting approaches, while reducing training time by over an order ofmagnitude.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×