Solving Interactive Video Object Segmentation with Label-Propagating Neural Networks

Interactive Video Object Segmentation (IVOS) addresses the problem of pixel-wise video annotation in collaboration with a human supervisor. Deep convolutional network-based approaches are a popular choice for solving IVOS problems. However, their performance is limited unless millions of parameters and huge datasets are used in their training. In this paper, we investigate an alternative solution, that interprets the core of the IVOS problem as a graph-based semi-supervised learning task and solves it with a label-propagating graph neural network architecture. The training of our method is sample efficient and outperforms all other methods with a similar property on the DAVIS dataset. We release the source code and the trained models.