Semester of Graduation

Spring 2022

Degree

Master of Science (MS)

Department

Electrical Engineering in Division of Electrical & Computer Engineering

Document Type

Thesis

Abstract

Segmentation is a process of partitioning a digital image or frame into multiple regions or objects. The goal of segmentation is to identify and locate the objects of interest with their boundaries. Recent segmentation approaches often follow such a pipeline: they first train the model on a collected dataset and then evaluate the trained model on a given image or video. They assume that the appearance of object is consistent in training and testing sets. However, the appearance of object may change in different photography conditions. How to effectively segment the objects with volatile appearance remains under-explored. In this work, we present a framework for image and video segmentation of appearance-volatile objects, including two novel modules, uncertain region refinement and feature bank. For image segmentation, we designed a new confidence loss and a fine-grained segmentation module to enhance the segmentation accuracy in uncertain regions.

For video segmentation, we proposed a matching-based algorithm which feature banks are created to store features for region matching and classification. We introduced an adaptive feature bank update scheme to dynamically absorb new features and discard obsolete features.

We compared our algorithm and the state-of-the-art methods on the public benchmarks. Our algorithm outperforms the existing methods and can produce more reliable and accurate segmentation results.

Committee Chair

Li, Xin

DOI

10.31390/gradschool_theses.5495

COinS