Study-unit DEEP LEARNING AND ROBOT PERCEPTION
| Course name | Computer engineering and robotics |
|---|---|
| Study-unit Code | A003169 |
| Curriculum | Robotics |
| Lecturer | Gabriele Costante |
| Lecturers |
|
| Hours |
|
| CFU | 9 |
| Course Regulation | Coorte 2023 |
| Supplied | 2024/25 |
| Supplied other course regulation | |
| Learning activities | Caratterizzante |
| Area | Ingegneria informatica |
| Sector | ING-INF/04 |
| Type of study-unit | Obbligatorio (Required) |
| Type of learning activities | Attività formativa monodisciplinare |
| Language of instruction | Italian |
| Contents | Image filtering, keypoint detection and feature extraction. Keypoint matching and Image Transformations. Epipolar geometry and stereo reconstruction. Visual odometry, graph-based SLAM and visual SLAM. Deep Learning and convolutional neural networks. Deep Reinforcement Learning. Guided exercises on Pytorch for the development of solutions based on convolutional neural networks and deep reinforcement learning. |
| Reference texts | Richard Szeliski. “Computer Vision: Algorithms and Applications”, Springer - Ian Goodfellow, Yoshua Bengio, Aaron Courville. “Deep Learning”, MIT Press - Christopher Bishop. “Pattern Recognition and Machine Learning”, Springer; Richard Hartley, Andrew Zisserman. “Multiple view geometry”, Cambridge University Press – Ramesh Jain, Rangachar Kasturi, Brian G. Schunk. “Machine Vision”, McGraw-Hill - Richard S. Sutton, Andrew G. Barto. “Reinforcement Learning: an Introduction”, MIT Press. Teacher's handouts available on UNISTUDIUM - PIATTAFORMA DI E-LEARNING DELL'UNIVERSITÀ DEGLI STUDI DI PERUGIA (https://www.unistudium.unipg.it/unistudium/ ) |
| Educational objectives | - Basic knowledge of: methodologies for the extraction of information of various semantic levels from images; techniques for the feature extraction and image descriptors computation; techniques for transformation estimation between pairs of images; strategies for 3D reconstruction from stereo images; principles and intuition behind Visual odometry and Visual SLAM; characteristics, advantages and solutions based on convolutional neural networks; methodologies and solution based on reinforcement learning and deep reinforcement learning. - Skills: capability to choose technologies and functional blocks of image analysis systems for various applications; capability to use convolutional neural networks and to design deep learning and deep reinforcement learning architectures (in particular, using the Pytorch framework) - The teaching also contributes to the following learning outcomes: capability to elaborate and/or apply original ideas in different applications; capability to solve problems in new and/or interdisciplinary environments; capability to motivate the design choices made, highlighting possible critical issues; capability to integrate knowledge from different sources and manage complexity. |
| Prerequisites | In order to understand the content presented and achieve the learning objectives it is useful to have a basic knowledge of Linear Algebra and a good knowledge of Machine Learning and programming. Suggested teaching: "Machine Learning and Data Mining". |
| Teaching methods | The teaching is organized as follows: - frontal lectures in the classroom; - seminars - guided exercises at the Computer Science Laboratory on the design of solutions for deep learning and deep reinforcement learning problems with the Pytorch framework. Teaching support tools: blackboard and PC+projector, PC. |
| Other information | I Semester (more details at link http: //www.ing.unipg.it/it/didattica/studiare-nei-nostri-corsi/orario-delle-lezioni ). |
| Learning verification modality | The verification of learning is performed with an oral exam and a practical test (project). The oral test includes a discussion of the submitted project and two questions on topics chosen by the teacher, and lasts about 45 minutes. The exam aims to verify the student's understanding of the fundamental theoretical tools provided by the teaching, as well as the relationship between these tools and the general topics of information engineering. The project aims to verify the student capabilities to design solutions based on deep learning and deep reinforcement learning methodologies for computer vision and robotic applications by using the Pytorch framework. The project needs to be sent to the teacher the day before the examination date. Reservations for examinations are made via the SOL portal: https://www.segreterie.unipg.it/. For information on support services for students with disabilities and/or DSA visit http://www.unipg.it/disabilita-e-dsa. |
| Extended program | Introduction to Computer Vision. Image Filtering Image Features. Feature Descriptor and Matching. Image Transforms. Stereo geometry. Visual odometry. Graph-based SLAM and Visual SLAM Machine learning and Computer Vision: Introduction. Deep Learning. Convolutional Neural Network (CNN). CNN Architectures. Recurrent Neural Network: RNN –LSTM –GRU. CNN Applications. Generative Models (Generative adversarial network (GAN)) and CNN unsupervised applications. Reinforcement Learning: Introduction, basic principles, and algorithms. Deep Reinforcement Learning. Laboratory sessions: framework for deep learning, CNN Architecture design, CNN application examples. Deep Reinforcement Learning application examples. |
| Obiettivi Agenda 2030 per lo sviluppo sostenibile |