Spatially Adaptive DGL Test for Robust User-Assisted Multilabel Segmentation

Abstract

Recently, the DGL test has been successfully applied to the user-assisted image segmentation problem where different types of user inputs, e.g. labeled pixels from ground truth masks, bounding boxes and pixel seeds, can be robustly leveraged to assist the segmentation process in a simple and effective way. However, in the baseline method the spatial information of the user inputs is not utilized and the test is implemented in the color domain. In this work, we propose a spatially adaptive version of the DGL test where the spatial information of the user-input regions is incorporated into the decision making process of the original test for an improved segmentation performance. We show that the proposed approach can be simply and seamlessly integrated into the baseline method without increasing its computational and algorithmic complexity. We demonstrate simulations on the Berkeley’s BSDS500 image database that validate the effectiveness of the proposed method. We also present benchmarking results which indicate that the accuracy can be improved by about 3% compared to the baseline method.