DEEP INTERACTIVE IMAGE SEGMENTATION

Abstract

Interactive image segmentation has been extensively studied in the past decade due to its many application domains. However, previous methods typically rely on low-level handcrafted features, which often fail in many challenging scenarios. Recently, the first deep learning-based interactive image segmentation method was proposed with significant improvements over previous techniques. Despite its excellent performance, it fails to fully utilize user-provided information and suffers from contextual ambiguity when the amount of user inputs is limited. In this thesis, we extend the deep learning-based approach to address the aforementioned challenges. Specifically, the proposed methods exploit local regional context surrounding the user inputs for local refinement, discover potential segmentation errors to serve as extra user inputs for self-refinement and generate a set of diverse segmentation masks that conform to the user inputs to reduce contextual ambiguity. They achieve state-of-the-art performance on various challenging benchmarks, demonstrating their effectiveness for practical applications.