RoI Tanh-polar transformer network for face parsing in the wild

Face parsing aims to predict pixel-wise labels for facial components of a target face in an image. Existing approaches usually crop the from input image with respect bounding box calculated during pre-processing, and thus can only parse inner Regions Interest (RoIs). Peripheral regions like hair are ignored nearby faces that partially included cause distractions. Moreover, these methods trained evaluated on near-frontal portrait images their performance in-the-wild cases has been unexplored. To address issues, this paper makes three contributions. First, we introduce iBugMask dataset wild, which consists 21,866 training 1000 testing images. The obtained by augmenting existing large poses. manually annotated 11 there variations sizes, poses, expressions background. Second, propose RoI Tanh-polar transform warps whole representation fixed ratio between area context, guided box. new contains all information original image, allows rotation equivariance convolutional neural networks (CNNs). Third, hybrid residual learning block, coined HybridBlock, layers both space Tanh-Cartesian space, allowing receptive fields different shapes CNNs. Through extensive experiments, show proposed method improves state-of-the-art wild does not require landmarks alignment.