Graduate Thesis Or Dissertation
 

Part-based and Uncertainty-Aware Few-shot Object Segmentation in Images

Public Deposited

Downloadable Content

Download PDF
Request Accessible Version
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/k643b801g

Descriptions

Attribute NameValues
Creator
Abstract
  • This dissertation addresses few-shot object segmentation in images. The goal of segmentation is to label every image pixel with a class of the object occupying that pixel, where the class may represent a semantic object category or instance. In few-shot segmentation, training and test datasets have different classes. Every new class in the test dataset is represented by a few support images and the corresponding query images, where we have access to the ground-truth segmentation of the support images, and the goal is to segment the corresponding queries. We focus on three related problems: (1) few-shot semantic segmentation; (2) few-shot instance segmentation; and (3) incremental few-shot instance segmentation. Our key technical contributions include an extension of Mask-RCNN -- the state-of-the-art instance segmenter aimed at general settings with sufficient training data -- to work in the few-shot setting. First, from the support images and their ground-truth segmentations, we extract more discriminative, diverse, and general feature vectors to represent the target class. Second, we develop a new object classifier that leverages Bayesian learning to take into account the paucity of labeled examples of testing classes. Third, we specify new loss functions for object bounding-box and segmentation-mask prediction, which take into account the uncertainties of prediction and use these uncertainties to regularize training. Finally, we introduce a new part-based instance segmenter that explicitly models latent object parts, which are shared by training classes and thus are expected to facilitate learning of testing classes on a few examples. Our empirical evaluation demonstrates the advantages of our contributions relative to existing work.
License
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Academic Affiliation
Rights Statement
Publisher
Peer Reviewed
Language
Accessibility Feature

Relationships

Parents:

This work has no parents.

In Collection:

Items

Thumbnail Title Date Uploaded Visibility Actions
file details NguyenKhoiD2021.pdf 2021-06-11 Public Download