2021/08/31 by Роман Шаповалов, David Novotný, Shapovalov, Roman +8 · 2 citations
Computer Science · #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Human Pose and Action Recognition #I.4.5 #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.2109.00033
openalex publication_date 2021/08/31 · openalex created_date 2022/10/01 · openalex updated_date 2026/07/28
We tackle the problem of monocular 3D reconstruction of articulated objects\nlike humans and animals. We contribute DensePose 3D, a method that can learn\nsuch reconstructions in a weakly supervised fashion from 2D image annotations\nonly. This is in stark contrast with previous deformable reconstruction methods\nthat use parametric models such as SMPL pre-trained on a large dataset of 3D\nobject scans. Because it does not require 3D scans, DensePose 3D can be used\nfor learning a wide range of articulated categories such as different animal\nspecies. The method learns, in an end-to-end fashion, a soft partition of a\ngiven category-specific 3D template mesh into rigid parts together with a\nmonocular reconstruction network that predicts the part motions such that they\nreproject correctly onto 2D DensePose-like surface annotations of the object.\nThe decomposition of the object into parts is regularized by expressing part\nassignments as a combination of the smooth eigenfunctions of the\nLaplace-Beltrami operator. We show significant improvements compared to\nstate-of-the-art non-rigid structure-from-motion baselines on both synthetic\nand real data on categories of humans and animals.\n