FI-BRS: A Few-Interaction Backpropagating Refinement Scheme for Video Object Segmentation

As a backpropagating refinement scheme (BRS) is introduced as an interactive algorithm to combine deep neural network and user-specified context for object segmentation, we further propose a Few-Interaction BRS (FI-BRS) for video Object segmentation (VOS) task. With only few user’s scribbles, FI-BRS automatically selects key regions (markers) inside and outside a target object in reference frame to play the role of scribbles. To form the distance maps in a BRS, the corresponding markers in test frame are located due to feature similarity with the ones in reference frame. For higher efficiency, the markers follow a parallel input order to form distance maps, which has a faster speed without obvious accuracy loss than the conventional way of serial input. The good performances on DAVIS and Youtube-Object benchmarks demonstrate that our proposed FI-BRS yields competitive results compared with state-of-the-art VOS methods.