Multi-Branch U-Net for Interactive Segmentation

In deep learning-based interactive segmentation, one of the key issues is how to effectively incorporate label prior into the network. The sparsity of clicks and the difference between images and label information limit the compatibility of user intent embedding in the deep long-path network. This letter aims to explore more effective embedding strategy of label prior in the neural network, and proposes a multi-branch U-net architecture for the interactive segmentation task. A principled fusion manner is adopted to progressively enhance the semantics of features based on the layer-by-layer promotion of the label prior guidance. Vast experiments on the popular GrabCut, Berkeley, DAVIS and SBD datasets verified the effectiveness of the proposed method.