Abstract
Accurate human shape recovery from a monocular RGB image is a challengingtask because humans come in different shapes and sizes and wear differentclothes. In this paper, we propose ShapeBoost, a new human shape recoveryframework that achieves pixel-level alignment even for rare body shapes andhigh accuracy for people wearing different types of clothes. Unlike previousapproaches that rely on the use of PCA-based shape coefficients, we adopt a newhuman shape parameterization that decomposes the human shape into bone lengthsand the mean width of each part slice. This part-based parameterizationtechnique achieves a balance between flexibility and validity using asemi-analytical shape reconstruction algorithm. Based on this newparameterization, a clothing-preserving data augmentation module is proposed togenerate realistic images with diverse body shapes and accurate annotations.Experimental results show that our method outperforms other state-of-the-artmethods in diverse body shape situations as well as in varied clothingsituations.