Deep Local Shapes: Learning Local SDF Priors for Detailed 3D Reconstruction

LEAP: Learning Articulated Occupancy of People

“…Even though the problem of the memory complexity for the grid-based methods is approached by [27,43,57,66,67], they have been outperformed by the recent learning-based continuous representations [2,3,10,12,19,30,39,40,42,45,46,56,62,64]. Furthermore, to improve scalability and representation power, the idea of using local features has been explored in [7,11,41,46,51,52,62]. These learning-based approaches represent 3D geometry by using a neural network to predict either the closest distance from a query point to the surface or an occupancy value (i.e.…”

“…where w k,i is an element of W. Specifically, let G = {G k (θ, J) ∈ R 4×4 } K k=1 be the set of K rigid bone transformation matrices that represent a 3D human body in a world coordinate (7). Then, B = {B k ∈ R 4×4 } K k=1 is the set of local bone transformation matrices that convert the body from the canonical space to a posed space (8), and j j ∈ R 3 (an element of J ∈ R K×3 ) represents jth joint location in the rest pose.…”

“…Overview. LEAP is an end-to-end differentiable occupancy function f Θ (x|G) : R 3 → R that predicts whether a query point x ∈ R 3 is located inside the 3D human body represented by a set of K rigid bone transformation matrices G (7). The overview of our method is depicted in Fig.…”

SCALE: Modeling Clothed Humans with a Surface Codec of Articulated Local Elements

“…Local Shape Modeling. Instead of learning 3D shapes solely with a global feature vector, recent work shows that learning from local shape variations leads to detailed and highly generalizable 3D reconstruction [11,13,22,52,55,64,70]. Leveraging local shape priors is effective for 3D reconstruction tasks from 3D point clouds [11,13,22,28,55,70] and images [52,64,76].…”

“…Instead of learning 3D shapes solely with a global feature vector, recent work shows that learning from local shape variations leads to detailed and highly generalizable 3D reconstruction [11,13,22,52,55,64,70]. Leveraging local shape priors is effective for 3D reconstruction tasks from 3D point clouds [11,13,22,28,55,70] and images [52,64,76]. Inspired by this prior work, SCALE leverages both local and global feature representations, which leads to high-fidelity reconstruction as well as robust generalization to unseen poses.…”