Majumder, Soumajit: Maximizing Information from User-Clicks for Efficient Instance Segmentation in Images. - Bonn, 2022. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5-67570
@phdthesis{handle:20.500.11811/10191,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5-67570,
author = {{Soumajit Majumder}},
title = {Maximizing Information from User-Clicks for Efficient Instance Segmentation in Images},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2022,
month = aug,

note = {Interactive instance segmentation allows users to select and obtain accurate pixel-level masks for objects of interest by providing inputs such as clicks, scribbles, or bounding boxes. It has always been a problem of interest in computer vision research, as it addresses quality problems faced by fully automated segmentation methods. The segmented results are helpful for downstream applications such as human-machine collaborative annotation, image/video editing, and mage-based medical diagnosis. The goal is to obtain accurate pixel-level masks for objects with minimal user input. In this dissertation, we propose several frameworks for performing interactive instance segmentation using user-provided clicks.
In interactive instance segmentation, users give feedback to refine segmentation masks iteratively. Typically, such frameworks refine false negatives and false positive regions via a succession of ‘positive’ and ‘negative’ clicks placed centrally in these regions. These user-provided ‘positive’ and ‘negative’ clicks are transformed into separate guidance maps that provide the network with necessary cues on the whereabouts of the object of interest. Most interactive frameworks incorporate these guidance maps at the image input layer. Our work proposes a novel transformation of user clicks to generate content-aware and location-aware guidance maps that leverage the hierarchical structural information present in an image. Using our guidance maps, even the most basic fully convolutional networks (FCNs) are able to outperform existing approaches that require state-of-the-art segmentation networks. Next, we propose an intuitive alternative for ‘positive’ and ‘negative’ refinement clicking by letting users click on the object boundary. We also propose a new multi-stage guidance framework for interactive segmentation. By incorporating user cues at different stages of the network, we allow user interactions to impact the final segmentation output more directly. We investigate and address challenges pertaining to user-click representation, refinement strategy, and network design in this work.
Through this dissertation, we advanced the state-of-the-art in interactive instance segmentation, proposed novel user click transformations and refinement strategies, presented new insights on the task-specialized design of such interactive frameworks. We demonstrated the effectiveness of our frameworks through comprehensive experimentation and by comparing them with existing state-of-the-art on standardized public benchmarks. We conclude this dissertation by presenting open challenges and outlining future research directions for interactive instance segmentation research.},

url = {https://hdl.handle.net/20.500.11811/10191}
}

The following license files are associated with this item:

InCopyright