T2: Review of State-of-The-Art Deep Learning Approaches for Visual Object Recognition and Tracking: Applications to Unmanned Aircraft Systems
Date and time: June 6, 2023, time to be announced (half day)
Organizers and Presenters
Mohammad H. Mahoor, Professor
D. F. Ritchie School of Engineering and Computer Science
University of Denver, Denver, CO, USA
Kimon P. Valavanis, Professor
D. F. Ritchie School of Engineering and Computer Science
University of Denver, Denver, CO, USA
Ali Pourramezan Fard, PhD Candidate
University of Denver, Denver, CO, USA
Jian Sun, PhD Candidate
University of Denver, Denver, CO, USA
Tutorial Summary
We witness an enormous proliferation of deep learning (DL) methods in robotics, intelligent systems, data science, and UAS, to say the least. Although Convolutional Neural Networks (CNNs) have been the ‘go-to’ architecture for Computer Vision tasks for the last decade, in the past two years, the Computer Vision community has witnessed massive growth in the number of new architectural designs.
As such, this tutorial centers around presenting and discussing deep learning (DL) algorithms, methods, approaches, and techniques, coupled with tools and support technologies for visual object recognition and object tracking, with a focus on unmanned aircraft systems (UAS). In detail, this tutorial focuses on three new classes of architecture: Transformer-based designs, Graph Neural Networks, and self-learning.
The topics to be covered are:
• Review of CNNs and state-of-the-art networks:
Brief review of CNN and well-known architectures for visual object detection, recognition, etc., with emphasis on applications related to UAS, such as real-time segmentation, thermal imaging, and behavior prediction.
• Emergence of new architectural designs:
Key recent innovations in architecture are presented, which deviate significantly from traditional CNNs. The discussion is on how transformer architectures bridge the gap between the vision domain and the natural language processing domain. This includes ViT and variants, Transformer-CNN hybrids, and MLP-based designs.
• Self-supervised learning (SSL):
SSL is an evolving machine learning technique poised to solve challenges posed by the over-dependence of labeled data. SSL obtains supervisory signals from the data itself, often leveraging the underlying structure of the data. Presentations center on covering the basics of self-supervised learning, the use of unlabeled data for pre-training purposes and ultimately improving classification accuracy.
• Graph Neural Networks:
Graph Machine Learning presents powerful tools to tackle this representation learning problem. Recent research has successfully applied Graph Machine Learning to handle several problems in the visual computing area, such as geometric processing, scene graph generation, video understanding, multi-object relational mining, visual navigation and so on. As such, this part reviews and discusses GNNs with applications in visual object recognition, object tracking, etc.
• State-of-the-art DL techniques and models applied in UAS:
The technical community has witnessed derivation of a plethora of techniques, algorithms and approaches related to modeling, navigation, and control of UAS. This part focuses on answering the following fundamental questions when it comes to applying DL to UAS: What do we learn? How do we learn? How fast do we learn? Do we learn offline, online, or both? Is the algorithm/technique suitable for (hard) real-time application? In addition, a summary of approaches tacking visual object recognition and tracking in UAS using novel networks (Vision Transformers, Graph models, self-learning etc.) is presented.
Intended Audience
This Tutorial is suitable for graduate students, researchers, scientists and engineers, practitioners, end-users, and developers interested in deep learning and computer vision with applications in autonomous UAVs. The collective outcome of the Tutorial is an understanding of new architectures in deep neural networks such as transformers, and graph neural networks, and new learning approaches such as self-supervised learning. We will also discuss how these methods and papers that have recently been applied in UAS.
Tutorial Material
Participants will receive detailed presentations and papers.