GIACOMO BORACCHI - TEACHING
|
||
|
Advanced Deep Learning Models and Methods for 3D Spatial Data AA 2023/2024, PhD Course, Politecnico di Milano Overview: This course presents recent advances in deep learning that brought data-driven models to achieve state-of-the-art performance in solving 3D vision problems. In particular, students will become acquainted with the biggest challenges of handling 3D data that are scattered in nature, thus are not suited for traditional filtering operations underpinning convolutional layers. The course will illustrate the most important layers for handling 3D data, as well as the neural networks for solving 3D Computer Vision problems and their application to Robotics and Computational Geometry. More information on the Course program page Organizers: Giacomo Boracchi, Matteo Matteucci. Politecnico di Milano. Dates: From January 17th 2024 to February 14th 2024. Schedule and Abstracts: Deep Learning for Volumetric Data and 3D Point Clouds (Boracchi): Giacomo Boracchi Professor at Politecnico di Milano January 17th, 14:15 - 18:00 Sala Conferenze, DEIB, Building 20 The most popular and successful deep learning models are meant for data lying over a rigid grid. Fully connected and convolutional layers are designed to process scattered data, as Point Clouds or 3D meshes returned by 3D sensors that are nowadays ubiquitous. In this lecture we will provide a formal description of 3D point clouds and meshes, and illustrate the type of sensors where these originates from. Then, we will introduce the mainstream solutions for solving visual recognition problems on Point Clouds (e.g. PointNet), including point-convolutional layers that extend the popular 2D convolutional layers to these type of data (KPConv). The most successful deep learning models for solving Point Cloud classification and semantic segmentation will be then presented. Lecture Slides , Colab Folder , Deep Leaerning for Depth Estimation: Luca Magri Researcher at Politecnico di Milano January 24th, 14:15 - 18:00 Sala Conferenze, DEIB, Building 20 Deep Learning for Depth Estimation: Obtaining dense and accurate depth measurement from images is a fundamental task for many 3D computer vision applications. In the last few years, depth estimation has undergone a paradigm shift due to the introduction of learning-based methods that have successfully replaced heuristics and hand-crafted rules. In particular, we will introduce the most relevant solution to single-view depth estimation which is an inherently ill-posed problem, and becomes tractable thanks to learned prior on 3D scenes. Then, we will move to multi-view depth estimation networks that improves the results even in configurations that are challenging for traditional methods (wide baseline) and exploits geometric constraints in the regression of the depth. Lecture Slides , Deep Learning for Non Rigid Shape Registration: Simone Melzi Researcher at Università Milano Bicocca January 31st, 14:15 - 18:00 Sala Conferenze, DEIB, Building 20 A challenging problem in computational geometry is to estimate correspondences between two shapes representing two deformed versions of the same entity (e.g., a dog in two different poses) or two entities from the same class (e.g., two chairs). This task, known as shape registration, finds several applications including texture/deformation transfer and statistical shape analysis. Many solutions for shape registration arise from rigid counterparts, by extending mainstream solutions like iterative closest points (ICP), coherent point drift (CPD), and methods based on the definition of pointwise descriptors (SHOT, HKS, among others). In the last decade, the functional approach (Functional maps) has given rise to a vast family of efficient alternatives. More recently, the large availability of data-driven solutions paved the way for machine learning procedures that rapidly outperformed all axiomatic competitors. In this lecture, we will overview some of the most impactful shape registration pipelines focusing in particular on the ones that inspired recent data-driven solutions (pointwise signatures and functional maps). Furthermore, we will analyse how well-known machine learning architectures have been applied to the shape registration task (CNN and transformers). Lecture Slides , Deep Learning in 3D for Robotics: Matteo Matteucci Professor at Politecnico di Milano January 17th, 14:15 - 18:00 Sala Conferenze, DEIB, Building 20 Leveraging on the background from the previous lectures applications of deep learning methods for 3D data processing in robotics and autonomous vehicles will be presented highlighting the challenges of interpreting the 3D semantics of the environments with the purpose of interacting with it. Examples of applications which will be presented concerns 3D object detection and 3D semantic scene parsing for autonomous driving, cooperative 3D perception and 3D loop detection. Lecture Slides , |
||