Geo-localization Framework for Real-world Scenarios - Defense Presentation
Offered By: University of Central Florida via YouTube
Course Description
Overview
Watch a 39-minute defense presentation by Sijie Zhu from the University of Central Florida on geo-localization frameworks. Explore the challenges of real-world scenarios, compare datasets, and learn about a novel geo-localization framework. Discover a new loss function that leverages multiple references and delve into orientation definition and estimation. Examine the predominant triplet-based loss and its adjustments for similarity distribution. Investigate methods to bridge the domain gap, including Vision Transformers and non-uniform cropping. Analyze retrieval performance on the VIGOR dataset, including meter-level evaluation and unknown orientation scenarios. Gain insights through visualizations and qualitative results presented in this comprehensive academic presentation.
Syllabus
Intro
Education Background
Overview
Toward Real-world Scenarios
Datasets Comparison
A Novel Geo-localization Framework
Novel Loss to Leverage Multiple Reference
Orientation Definition
Revisiting the Orientation Issue
The Predominant Triplet-based Loss
Better Adjustment on Similarity Distribution
Estimate the Orientation
Better Visual Explanation and Orientation Estimatio
How to Bridge the Domain Gap?
Vision Transformer (VIT)
Non-uniform Cropping
Retrieval Performance on VIGOR
Meter-level Evaluation
Unknown Orientation
Visualization
Qualitative Results-VIGOR
Taught by
UCF CRCV
Tags
Related Courses
Introduction to Deep LearningMassachusetts Institute of Technology via YouTube Taming Dataset Bias via Domain Adaptation
Alexander Amini via YouTube Making Our Models Robust to Changing Visual Environments
Andreas Geiger via YouTube Learning Compact Representation with Less Labeled Data from Sensors
tinyML via YouTube One-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning
University of Central Florida via YouTube