Date: September 2018.
Source: 15th European Conference on Computer Vision (ECCV 2018), Munich, Germany.
Research Summary: Realistic cloth modeling has traditionally relied on either intensive physics-based simulation, which requires numerous heuristic parameters, or models reconstructed from visual observations, which typically lack fine geometric detail. DeepWrinkles addresses this gap with a framework combining two modules that work jointly to represent both global shape deformation and fine surface detail with high fidelity. Global shape deformations are recovered from a subspace model learned from real-world 3D data of clothed people in motion, while high-frequency wrinkle detail is added through normal maps generated by a conditional Generative Adversarial Network (GAN) designed to enforce realism and temporal consistency across a sequence. Because the model is learned independently from body shape and pose, it supports applications requiring retargeting, such as body animation. The result is an entirely data-driven approach to realistic cloth wrinkle generation, achieving high-quality rendering of clothing deformation sequences from real, high-resolution observations.
3dMD’s Role: DeepWrinkles’ clothing deformation sequences were captured at 60 FPS using a dynamic-3D/4D 3dMDbody system, producing colored meshes of 200,000 vertices per frame. The DeepWrinkles model was trained on 9,213 consecutive 3dMD frames of real clothing movement, split into training, test, and validation sets.
Article: DeepWrinkles: Accurate and Realistic Clothing Modeling.
Authors: Zorah Lähner, Daniel Cremers, Tony Tung. Meta (Facebook) Reality Labs and Technical University Munich.