Search Results for "surface reconstruction"

Mesh R-CNN is a 3D reconstruction and object understanding framework developed by Facebook Research that extends Mask R-CNN into the 3D domain. Built on top of Detectron2 and PyTorch3D, Mesh R-CNN enables end-to-end 3D mesh prediction directly from single RGB images. The model learns to detect, segment, and reconstruct detailed 3D mesh representations of objects in natural images, bridging the gap between 2D perception and 3D understanding.

...It first checks whether the image is suitable, then creates an ObjectSculptSpec describing geometry, materials, hierarchy, pivots, sockets, and visual priorities. Reconstruction proceeds through blockout, form, look development, and interaction phases so details are added in a controlled order. The plugin supports hard-surface objects, organic forms, vegetation, fabrics, glass, hair, emissive surfaces, and decals. Generated models can include animation-ready parent-child relationships and detachable or articulated parts. ...

TRELLIS.2 is a cutting-edge open-source model and codebase for high-fidelity 3D asset generation from 2D images, developed to push forward the state of the art in image-to-3D generation. At its core is a novel sparse voxel structure called O-Voxel that jointly encodes both geometry and surface appearance, enabling reconstruction and generation of complex 3D shapes with arbitrary topology, open surfaces, and physically based rendering (PBR) textures. The system leverages a large 4-billion-parameter architecture combining sparse 3D variational autoencoders with flow-matching transformers to produce fully textured 3D models at resolutions up to 1536³ voxels. ...

DeepSDF is a deep learning framework for continuous 3D shape representation using Signed Distance Functions (SDFs), as presented in the CVPR 2019 paper DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation by Park et al. The framework learns a continuous implicit function that maps 3D coordinates to their corresponding signed distances from object surfaces, allowing compact, high-fidelity shape modeling. Unlike traditional discrete voxel grids or meshes, DeepSDF...