Semantic 3D Reconstruction of Heads

i x i s = 1, x i s ≥ 0, x ij s ≥ 0 . • x i s ∈ [0, 1]: indicates whether label i is chosen at voxel s • y ij s = x ij s − x s ∈ [−1, 1]: presence of a surface and its direction • ρs: unary term, cost for label i at location s •φij s : convex smoothness term at voxel s for labels i and j • ek ∈ R: k-th canonical basis vector • T : alignment between prior and input data •α: normalization factor for scale •Convex in x, not convex in T (optimized using alternating optimization)