Interactive Segmentation Using Prior Knowledge-Based Distance Map

Abstract

In this paper, we aim to solve problems in interactive segmentation, a technique which is widely used for data labeling tasks. It requires the user to provide clicks for the objects of interest. The user-provided clicks are transformed into the distance map, which plays an important role in the interactive segmentation. Therefore, we propose a novel distance map that is obtained by combining the automatic segmentation result with the user-provided clicks. Since we have validated that better automatic segmentation result leads to better interactive segmentation result, we concatenate the original image with its LOG (Laplacian of Gaussian) filter image to improve the automatic segmentation results. Besides, given that its successful implementation requires correct labels so as to enable the computer to simulate the user interaction, a data cleansing technique is applied to filter out samples with inaccurate labels also known as noisy labels. The effectiveness of our proposed method is assessed using the Kaggle’s TGS Salt Identification Challenge dataset. The obtained results indicate that when using the proposed algorithm, the average IoU reaches 91.81% for only one user-provided click.