Supplementary Materials for Regional Interactive Segmentation Networks

The Hamming distance is commonly used in the segmentation literature to measure the pixel-wise difference between two binary maps where 0 implies the two completely overlap. Here, we employ the Hamming distance metric to measure both the pixel-level accuracy and confidence of each pixel (close to 0 or 1). The smaller the Hamming distance, the more accurate is the model. In addition, we also compute the coverage score to measure the recall rate of each model where the model with higher coverage score is preferred. Figure 1 shows an intuitive example for interpreting these two metrics. The results on Grabcut and Berkeley dataset are shown in Figure 2.