Fusing partitions of diverse hand-crafted features derived from unsupervised intermediaries for robust and interpretable object co-associations
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
Data association in digital imagery is a complex task in any setting, but remote sensing provides a uniquely challenging context. Indeed, rapidly changing viewpoints, varying sensor modalities, and atmospheric distortions introduce many obstacles to creating reliable methods for processing remotely sensed imagery (RSI). While ultimately concerned with object tracking in RSI, this work first explores general methods for extracting and ensembling features in multi-modal contexts. A central contribution here involves an explainable, confidence-based algorithm for leveraging ensembles of weak feature associators to perform multi-feature data association. This approach gives humans the ability to understand the decisions of and further fine-tune the association algorithm that underpins downstream tasks (e.g., object tracking, image retrieval, or cohort discovery). Highly parallelizable and modular by design, new features can be introduced, removed, suppressed, or prioritized without drastic impacts on association computation time. Additionally, the proposed recursive lifetime search provides a new approach to extracting a final partition from the dendrogram created through agglomerative clustering.
Table of Contents
PubMed ID
Degree
Ph. D.