| Abstract: | In this paper, we propose a coarse-to-fine convolutional network framework designed with problem specific knowledge for fast automatic portrait segmentation. We built up a dataset of 7 100 portrait images which are frames from personal live show videos. The proposed network includes a coarse network which can learn global information and a fine network which utilizes local information to refine the coarse output. Additionally, an auxiliary contour loss is introduced to help training the coarse network. The proposed framework shows higher accuracy than the widely-used fully convolutional network. With light-weight post-processing, the predicted foreground mask can be used in real-time portrait video editing tasks such as background replacement. |
