A consistent ranking-based framework for segmentation

Download the PDF of the paper entitled “RankSEG: A Consistent Rank-based Framework for Segmentation” by Ben Dai and Chunlin Li.

Download PDF

overview: Segmentation has emerged as a fundamental area of ​​computer vision and natural language processing to assign labels to every pixel/feature to extract regions of interest from images/text. To assess segmentation performance, use the Dice and IoU metrics to measure the degree of overlap between ground truth and predicted segmentation. In this paper, we establish a theoretical basis for segmentation on Dyce/IoU metrics, including Bayesian rule and Dyce/IoU calibration, analogous to classification calibration and Fisher consistency in classification. Existing threshold-based frameworks with most operating losses prove to be inconsistent in terms of Dice/IoU metrics and thus can lead to sub-optimal solutions. To address this pitfall, we propose a new consistent ranking-based framework, namely RankDice/RankIoU, inspired by plugin rules for Bayesian segmentation rules. Three numerical algorithms with GPU parallel execution were developed to implement the proposed framework with large-scale and high-dimensional segmentation. We study the statistical properties of the proposed framework. We show that this is adjusted for Dice/IoU, and its excess risk bounds and convergence rate are also provided. The numerical effectiveness of RankDice/mRankDice is demonstrated on various simulated examples with state-of-the-art deep learning architectures and finely annotated CityScapes, Pascal VOC, and Kvasir-SEG datasets .