Please use this identifier to cite or link to this item: https://hdl.handle.net/2440/136542
Type: Thesis
Title: Deep Object Segmentation and Beyond
Author: Wang, Xinlong
Issue Date: 2022
School/Discipline: School of Computer Science
Abstract: Object segmentation is a fundamental computer vision problem that aims to recognize the object of interest and group the corresponding pixels in an image. With wide applications in self-driving cars, medical imaging, augmented reality, etc., object segmentation has attracted a lot of research attention. In this thesis, we propose a series of methods to solve the challenging problem with deep neural networks. We further generalize the proposed methods to solve extensive tasks such as image matting and study the interactions between object segmentation and unsupervised learning. First, we propose segmenting objects by locations (SOLO), a new, embarrassingly simple approach to segment all the object instances in an image and recognize their categories. Unlike previous methods that rely on either bounding box detection or grouping post-processing, SOLO directly maps a raw input image to the desired object categories and instance masks with a fully convolutional network. We demonstrate a much simpler and flexible instance segmentation framework with strong performance. Second, we present SOLOv2, a dynamic and fast instance segmentation solution that follows the principle of SOLO but improves it in terms of both speed and accuracy. SOLOv2 achieves state-of-the-art results with high efficiency, making it suitable for both mobile and cloud applications. We further demonstrate the generality of our method by extending it to perform panoptic segmentation and image matting. Third, we propose dense contrastive learning (DenseCL) to learn better representation from large-scale unlabeled images for dense prediction tasks such as segmentation. The proposed DenseCL performs dense pairwise contrastive learning at the level of pixels. Our method largely closes the gap between self-supervised pre-training and downstream dense prediction tasks. Finally, we propose a fully unsupervised learning method that learns to segment objects without any annotations. We present FreeSOLO, a self-supervised instance segmentation framework built on top of our simple-yet-effective methods SOLO(v2) for segmentation, and DenesCL for unsupervised learning. For the first time, we demonstrate unsupervised instance segmentation successfully. The code and models are publicly available at https://github.com/WXinlong.
Advisor: Shen, Chunhua
Dissertation Note: Thesis (Ph.D.) -- University of Adelaide, School of Computer Science, 2022
Keywords: nstance segmentation, object detection, panoptic segmentation, image matting, self-supervised learning
Provenance: This electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at: http://www.adelaide.edu.au/legals
Appears in Collections:Research Theses

Files in This Item:
File Description SizeFormat 
Wang2022_PhD.pdf17.32 MBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.