Watch a 40-minute thesis defense presentation by Yang Yang at the University of Central Florida exploring advanced computing features in computer vision. Delve into hand-designed features, feature learning problems, and a proposed method for event detection using learned filters and data-driven concepts. Examine the hybrid learning model, comparing it to generative and discriminative approaches. Explore applications in human and horse detection, and gain insights into gated auto-encoders for modeling relationships between videos. Conclude with a discussion on K-shot learning, composite datasets, and computational cost comparisons in this comprehensive exploration of cutting-edge computer vision techniques.
Thesis Defense: Computing Features in Computer Vision for Event Detection
University of Central Florida via YouTube
Overview
Syllabus
Intro
Computing features in computer vision
Hand-Designed Features
Feature learning problem
Outline
Common approach & Challenges
62 action concepts
Proposed Method
Event Detection Process
Learned Filters
3D Motion Filters
Data-driven Low level Features
Data Driven Concept (2D Scene)
Data Driven Concept (Motion)
Summary
The Model
Hybrid Learning
Experimental Setup
Hybrid Features
Hybrid Vs. Generative
Hybrid Vs. Discriminative
Higher Level Visualization
Human Detection Results
Performance on Horse Detection
Introduction
Flow Chart
Gated Auto Encoders - Model the relationship of two videos
Discriminative Learning
Pair of Features
Generative V.S. Discriminative
K-shot Learning
Composite dataset
Computational Cost Comparison
Conclusion
Taught by
UCF CRCV
