Real-Time Semantic Clothing Segmentation

Clothing segmentation is a challenging field of research which is rapidly gaining attention. This paper presents a system for semantic segmentation of primarily monochromatic clothing and printed/stitched textures in single images or live video. This is especially appealing to emerging augmented reality applications such as retexturing sports players’ shirts with localized adverts or statistics in TV/internet broadcasting. We initialise points on the upper body clothing by body fiducials rather than by applying distance metrics to a detected face. This helps prevent segmentation of the skin rather than clothing. We take advantage of hue and intensity histograms incorporating spatial priors to develop an efficient segmentation method. Evaluated against ground truth on a dataset of 100 people, mostly in groups, the accuracy has an average F-score of 0.97 with an approach which can be over 88% more efficient than the state of the art.