Key Techniques for Digital Human Modeling: Facial Portrait Editing, Eyelash Segmentation, and Real‑Time Loose Clothing Animation

The presentation introduces three major research directions in digital human technology. First, facial portrait editing techniques are described, including video‑based fat‑slim adjustment presented at ACM Multimedia 2021, double‑chin removal demonstrated at SIGGRAPH 2021, and hair removal methods that improve 3D reconstruction by eliminating hair interference.

Second, a dedicated eyelash segmentation dataset (EyelashNet) and baseline methods are detailed. The authors explain the challenges of eyelash handling in high‑precision 3D face reconstruction, describe a fluorescence‑based data‑capture system with 16 cameras and UV lighting, and outline a multi‑stage pipeline: data collection, mask generation with a segmentation network, iterative refinement using a synthetic pre‑training step (RenderEyelashNet), and final training of a GCA network for mask inference.

Experimental results show that the proposed pipeline outperforms existing methods both qualitatively and quantitatively, with extensive ablation studies confirming the contribution of each component.

Third, the paper presents a deep‑learning approach for real‑time animation of loose clothing (SIGGRAPH 2022). The method introduces virtual skeletons to model low‑frequency deformations and a motion network that predicts high‑frequency details. A radial‑basis‑function (RBF) network handles variable physical simulation parameters, enabling the system to generalize across different garments and motions.

Data preparation involves generating ground‑truth meshes with Houdini Vellum Solver, applying skin decomposition to obtain virtual bones, and training separate low‑frequency (GRU‑based) and high‑frequency (GRU + GNN + MLP) modules. Results demonstrate close visual similarity to ground truth, robust performance under large motions, and superior quantitative metrics (RMSE, STED) compared to prior work.

The authors conclude with future directions, emphasizing the importance of high‑fidelity face and clothing reconstruction for games, virtual humans, and the metaverse, and suggest low‑cost methods for real‑time, photorealistic digital humans.