CVML Programming Short Course and Workshop on Deep Learning and Computer Vision

Expected Outcome

Part A lectures will provide a solid background on the topics of Deep neural networks and DNN programming tools (e.g., PyTorch). Two programming workshops will take place.

1. Deep neural networks – Convolutional NNs: From multi-layer Perceptrons to deep architectures. Fully connected layers. Convolutional layers. Tensors and mathematical formulations. Pooling. Training convolutional NNs. Initialization. Data augmentation. Batch Normalization. Dropout. Deployment on embedded systems. Lightweight deep learning. DNN programming tools (e.g., PyTorch).

2. Knowledge Distillation in Deep Neural Networks: The course about Knowledge distillation in deep neural networks will provide participants with an in-depth understanding of the concept of distilling knowledge from a neural network. The lecture will cover the basics of knowledge distillation including its definition, applications and benefits. Participants will learn about the different techniques used, such as adversarial distillation, multi-teacher distillation, cross-modal distillation and others. In the hands-on section of the course, the participants will be able to apply and compare various distillation techniques in computer vision tasks.

Part B lectures will focus on computer vision algorithms, namely on 2D target tracking, Deep learning for object/face detection. Two programming workshops will take place.

1. Deep learning for object/face detection: Recently, Convolutional Neural Networks (CNNs) and Transformers have been used for object/target (e.g., car, pedestrian, road sign) detection with great results. However, using such models on embedded processors for real-time processing is prohibited by HW constraints. In that sense various architectures and settings will be examined in order to facilitate and accelerate the use of embedded CNN/Transformer-based object detectors with limited computational capabilities. The following object detection topics will be presented: Object Detection as classification and regression task. CNN-based architectures for object detection (e.g., RCNN, Faster-RCNN, YOLO, SSD) and Transformer-based ones (DETR). Lightweight architectures. Data augmentation. Deployment. Evaluation and benchmarking.

2. 2D target tracking: The 2D Object Tracking in Embedded Systems lecture offers a comprehensive exploration of object tracking techniques specifically tailored for embedded systems. This lecture delves into the intricacies of tracking objects in real-time within resource-constrained environments, where considerations such as limited processing power, memory, and energy consumption play crucial roles. Attendees will gain a deep understanding of optimized algorithms, sensor integration methods, and hardware acceleration techniques designed to achieve efficient and accurate object tracking in embedded systems. Practical examples and case studies with OpenCV programming will be presented to illustrate the application of these techniques in diverse domains, including robotics, surveillance, and autonomous vehicles, empowering attendees to implement robust and reliable object tracking solutions in their own embedded systems projects.

Part C lectures will offer the fundamentals of  Real-Time Image Segmentation and Image Segmentation on Natural Disaster Optical Flow data (e.g., videos of floods). Two programming workshops will take place.

1. Real-Time Image Segmentation: This course provides an overview of real-time image segmentation, focusing on its advancements including deep learning-based methods (CNNs and Transformers), its challenges such as limited computational resources and the need for high accuracy, and its applications. In this lecture, the fundamental concepts of image segmentation will be discussed, with emphasis on its significance in computer vision tasks in a natural disaster setting. In the hands-on section of the course, the participants will be able to apply various deep learning algorithms for real-time image segmentation in computer vision tasks related to Natural Disasters.

2. Natural Language Processing for Natural Disaster Management: The utilization of Natural Language Processing (NLP) in the domain of natural disaster management presents an opportunity for transformative advancements. This lecture will explore how we can use machine learning in NLP to improve important aspects of disaster response, such as early warning systems, situation awareness and emergency response coordination. The lecture will explore how these advanced algorithms facilitate the extraction of actionable intelligence from unstructured textual data sources, like social media feeds. Additionally, the session will highlight the power of NLP in enabling seamless communication and information sharing among response teams, aid organizations and affected communities.

Certification

Upon completion of the course, participants will be awarded a Certificate of attendance.

For the successful completion of the programme, the participants should:

1. have complete the questionnaires with mark at least 5/10.
2. to have paid all the tuition fees by 30/12/2023.