Semantic Segmentation of Fashion Photos using Light-Weight Asymmetric U-Net

Semantic segmentation is crucial for machine image understanding. The availability of public data set such as MSCOCO, ADE20K and CityScape encourages the development of popular models for semantic segmentation like SegNet and PSPNet. In this paper, we propose a light-weight deep neural network for street-fashion semantic segmentation. Experiment on ModaNet data set shows that our proposed network results in high accuracy despite its low requirement in the computational resource.