Hand over face segmentation using MPSPNet
Sakher Ghanem, Alex Dillhoff, Ashiq Imran, Vassilis Athitsos · 2020
Accurate hand segmentation is vital in many applications where the hands play a central role. Examples include sign language recognition, action recognition, and gesture recognition. A relatively unexplored obstacle to correct hand segmentation is the case of hands overlapping with the face. Inspired by the hand over face segmentation problem, we have developed the novel Multi-level Pyramid Scene Parsing Network (MPSPNet) for semantic segmentation. We evaluate MPSPNet on two standard object segmentation datasets (NYUDv2, PASCAL VOC) and two recently published and challenging datasets focusing on scenarios in which the hands overlap the face, VLM-HandOverFace and HOF. Additionally, we compare our method against several state-of-the-art hand segmentation methods such as RefineNet and PSPNet. We empirically show that the proposed method achieves a 6% improvement in mIOU compared with RefineNet on the VLM-HandOverFace dataset and a 15% improvement in mIOU compared with PSPNet on the HOF dataset.