Human parsing by weak structural label

Human parsing, which decomposes a human centric image into several semantic labels, e.g., face, skin etc, is an active topic in recent years. Traditional human parsing methods are always conducted on a supervised setting, i.e., the pixel-wise labels are available during the training process, which require tedious human labeling efforts. In this paper, we propose a weakly supervised deep parsing method to alleviate the human from the time-consuming labeling. More specifically, we resort to train a robust human parser with the structural image-level labels, e.g., “red jeans” etc. The structural label contains an attribute, e.g., “red”, as well as a class label, e.g., “jeans”. Our framework is based on the Fully Convolution Network (FCN) (Pathak et al. 2014) with two critical differences. First, the loss function defined on the pixel by FCN (Pathak et al. 2014) is modified to the image-level loss by aggregating the pixel-wise prediction of the whole image into a multiple instance learning manner. Besides, we develop a novel logistic pooling layer to constrain that the pixels responding to the color and corresponding category labels are the same to interpret the structural label. Extensive experiments in the publicly available dataset (Liu et al. IEEE Trans Multimedia 16(1):253–265, 2014) show the effectiveness of the proposed method.