Implementation and Analysis on Backpropagating Refinement Scheme for Interactive Image Segmentation

BRS is the first CNN-based interactive image segmentation algorithm to refine segmentation results based on a backpropagation scheme. In this paper, we give a detailed description of how BRS operates and demonstrate how to implement the algorithm. In BRS, user-provided clicks are first converted into interaction maps, which are then concatenated with the RGB image and provided as input to a segmentation network. In the test phase, performing the forward pass in the network generates an initial segmentation map. However, the user-annotated pixels may be mislabeled in the initial result. BRS refines this result by correcting the mislabeled pixels. We implement this BRS algorithm in PyTorch and publish the source codes. Moreover, we first show that BRS can reach the perfect IoU ratio of 1.0 in most cases and delineate objects more accurately than a variant of BRS, called f-BRS.