Postprocessing Method for Boundary-Preserving Edge Refinement for 3D Human Body Segmentation Sequences without Ground-Truth Data

We propose a post-processing strategy to improve the temporal stability of automatic human-body segmentation in sequences of 3D point clouds and polygonal meshes. While state-of-the-art segmentation methods—predominantly machine-learning based—are typically evaluated frame-by-frame, their direct application to time-ordered data often produces temporally inconsistent labels, manifested as boundary flickering between adjacent anatomical regions despite minor inter-frame motion. To address this limitation, we introduce a label reassignment procedure that refines vertex-/point-wise predictions by jointly exploiting temporal coherence from neighboring frames and spatial consistency within local geometric neighborhoods. The method corrects isolated misclassifications, suppresses outliers, and reduces spurious label transitions, yielding smoother and more anatomically consistent segment boundaries over time. Performance is quantified using six complementary cluster similarity measures—centroid, Hausdorff, Chamfer, Wasserstein, and Mahalanobis distances, together with anisotropy difference—computed over 14,500 samples. All meshes considered in this study were uniformly sampled, with an average inter-vertex spacing of approximately 15 millimeters. Results show that improvements arise when incorporating more than one adjacent frame and using denser spatial support, with the strongest gains achieved when the neighborhood’s mean spatial extent is approximately twice the vertex spacing (≈30 mm): mean reductions exceed ∼50% for: centroid (from more than 12 mm to ∼6 mm), Hausdorff (from over 40 mm to ∼20 mm), Wasserstein (from around 18 mm to 9 mm), and anisotropy, reach ∼30% for Chamfer (from 9 mm to just over 6 mm), and amount to ∼20% for Mahalanobis (from 2.30 to 1.85 mm). These reductions correlate with visibly smoother 3D playback and more stable inter-segment edges.