Gaussian correction for adversarial learning of boundaries

Social networking sites often monitor the response to brands, events and activities during personal chats or videos. Here, the facial expression of the speaker can be used for automatic ranking of products. However, manual classification of videos puts the identity of the speaker at risk. There is imminent danger of fake videos circulating that are generated using style transfer. In this paper, we target both these challenges by using an adversarial model that can segment a face from the background scenery and occlusions. The segmentation for a fake video will be of poor quality compared to a real video. Previous segmentation models could only be trained on a few objects and failed on scenic images with occlusions. Here we propose an image translator that learns the boundaries of objects during training using Gaussian correction. To determine the parameters of the Gaussian distribution we make use of a Lyapunov candidate function that converges to a global maximum. We apply the model to segmentation of faces and cars in photos. We also apply it to the task of style transfer to the background without affecting the foreground object. The proposed method outperforms baselines by over 20% on segmentation metrics such as IoU and BFScore.