National Style Representation Learning for Single-view 3D Garment Reconstruction

To address the problem of incomplete structure, inaccurate style and fuzzy local feature caused by the diversity and complexity of styles and accessories for minority clothing in single-view three-dimensional garment reconstruction, a national style representation learning method is proposed to learn and map the underlying style feature. Firstly, the shape underlying feature is learned by constructed shape representation using the defined shape style and geometric-topology of minority clothing. Secondly, the style representation is conducted based on regional location and key points by fusing the defined clothing style and dressed parts to obtain the local perception region maps. Then, combining the shape feature, the style feature and the defined symmetric loss function to implicitly reconstruct the preliminary model. Finally, the superpixel feature of image, a Branch network and the semantic parsing of accessories are added to the basis of convolutional network to establish accessory representation by encoding UV position map to generate the final model. The experimental results on minority garment dataset show that the chamfer distance and normal cosine distance error are 1.732 and 0.13 respectively, which reduce by 11% and 18%. The proposed method can improve the accuracy of three-dimensional national clothing reconstruction, which generates three-dimensional garment model with national style.