Temporally coherent person matting trained on fake-motion dataset

We propose a novel neural-network-based method to perform matting of videos depicting people that does not require additional user input such as trimaps. Our architecture achieves temporal stability the resulting alpha mattes by using motion-estimation-based smoothing image-segmentation algorithm outputs, combined with convolutional-LSTM modules on U-Net skip connections. also fake-motion generates training clips for video-matting network given photos ground-truth and background videos. apply random motion their simulate movement one would find in real composite result clips. It lets us train deep neural operating an absence large annotated video dataset provides training-clip foreground optical flow use loss functions.