Person Part Segmentation based on Weak Supervision

In this paper we address the task of semantic segmentation and propose a weakly supervised scheme for person part segmentation. Previous work on segmentation suffers from the lack in supervision due to the expensive labelling process. To address this issue, the correlation between depth estimation and semantic segmentation is explored in this paper and a scheme is proposed to utilize the knowledge learned from depth labels to serve the segmentation task. The supervision in the form of depth maps which can be acquired automatically contributes to the training of the segmentation model. The combination of the two types of supervision allows us to augment the training data of person part segmentation without additional annotations. Quantitative results have shown that the proposed scheme outperforms existing methods by over 5%. Qualitative results have also shown that the proposed scheme significantly outperforms existing methods.