STITCH: Structure Tensor Informed Topologically Coherent Hair Reconstruction
Year: 2026
Type: article
Abstract: Hair reconstruction remains fundamentally challenging due to the filamentary nature of strands, strong self‐occlusion, and the difficulty of recovering globally consistent 1D structure from multi‐view observations. While multi‐view cues provide complementary geometric evidence, existing approaches based on learned priors or implicit volumetric representations often struggle to generalize across diverse hairstyles or preserve fine directional detail. Learned methods can over‐constrain global structure due to limited training diversity, while implicit orientation fields tend to collapse multi‐modal strand orientations into locally averaged estimates, losing high‐frequency directional variation in heavily occluded regions. We depart from these approaches by introducing a sparse hierarchical reconstruction framework based on an ambiguous structure tensor representation that encodes local orientation distributions without committing to explicit strand directions. Our key insight is that preserving directional ambiguity during reconstruction enables robust aggregation of multi‐view evidence, postponing directional commitment until sufficient global information is available to enforce consistency. To reconstruct heavily occluded regions, we introduce a boundary‐driven propagation strategy that transfers reliable surface structure into the interior through locally coherent helical continuation, enabling unobserved geometry to remain consistent with surrounding strand flow. We further introduce a consensus‐based disambiguation mechanism that resolves tensor ambiguity and helical chirality into globally consistent strand growth directions, together with a cross‐view orientation regularization stage that stabilizes early surface reconstruction into thin, strand‐aligned structures. Extensive experiments demonstrate that our method significantly improves strand‐level fidelity and structural coherence on highly occluded, curly, and geometrically complex hairstyles.
Source: Computer Graphics Forum
DOI: 10.1111/cgf.70640
Language: English
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,710
Awards: G4866825924, G9001368724
OpenAlex ID: W7226019939
