Defining and Assessing a Novel Spatio-Chromatic Basis for Unsupervised, Class-Agnostic Scene Segmentation

Date

2021-08

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

This thesis defines a novel image feature, the "semantic pattern region", measures the strength of its correlation to scene semantics, and demonstrates its use for unsupervised, class-agnostic scene segmentation. The semantic pattern region is derived from spatio-chromatic image partitions and has an inherent one-to-one relationship with the semantic regions of the original image. I will demonstrate the pattern region's potential utility for segmentation in two ways: 1) by presenting representative examples of the best and worst results produced by the segmentation algorithm, and 2) by reporting the results of a quantitative assessment confirming a strong one-to-one spatio-chromatic correspondence between pattern regions and semantic and salient ground truth regions.

Description

Keywords

Segmentation Unsupervised

Citation