TL;DR
Given a single 2D image, unPIC reconstructs 3D objects via a shape-first hierarchical pipeline: it first predicts canonical 3D Normalized Object Coordinates (NOCS) in [0, 1]³ to establish global 3D geometry, then synthesizes appearance (RGB) conditioned on those 3D coordinates to unproject a dense 3D RGB point cloud and reconstruct a Screened Poisson 3D mesh.
Animals & Figurines
013D Shape Prior (NOCS)
(R, G, B) = (X, Y, Z) ∈ [0, 1]³
Loading 3D...
023D RGB Point Cloud
92,779 points
Loading 3D...
03Screened Poisson Mesh
102,771 triangles
Loading 3D...
2D → 3D Unproject
Point Size
All 40 Reconstructions
Showing 40 objects • Click any object to inspect in 3D
Baseline Comparison
Chamfer Distance (× 10⁻³, lower is better) evaluated on 8,192 points against Ground Truth mesh.