Neural rendering and reconstruction form the foundation of digitizing the physical world, with applications in virtual/augmented reality (VR/AR), film production, robotics, and beyond. This course explores advanced topics in this space, emphasizing data-driven approaches using neural models. After reviewing the basics of computer graphics and machine learning, we will delve into 3D scene representation and reconstruction using differentiable rendering, generative models, morphable models, human reconstruction and synthesis, and quantum visual computing.
Computer Vision strives to develop algorithms for understanding, interpreting and reconstructing information about real-world scenes from image and video data. Computer Graphics focuses on image synthesis: algorithms to build and edit static and dynamic virtual worlds and to display them in photorealistic or stylized ways. Both fields have witnessed the transformative effects of deep learning and neural networks, thus ushering the so-called Neural Age. In this seminar, we will explore how the classical concepts in Computer Vision and Computer Graphics have manifested in the Neural Age. We will seek to understand if and how the new neural methods have changed the way we think about these problems, and how they have improved the state-of-the-art. For each topic that we cover, we will discuss a seminal paper that has shaped the field and also discuss a recent paper that has further developed that idea in a modern context. This seminar will cover research papers from the following topics:
Neural rendering and reconstruction is the cornerstone of digitizing our world, with several applications in VR/AR, movie and media production, robotics, and many more. The digitization pipeline usually consists of three main stages; data capture, 3D model building, and finally, reconstruction and rendering/synthesis. This course will cover advanced topics in this digitization pipeline, with a focus on data-driven approaches using neural-based formulations. We will cover 3D scene representations, including explicit approaches as well as the more recent learnable implicit-based approaches. We will discuss how to build 3D morphable models for important objects such as the human face and body. We will also cover approaches for both 2D and 3D neural rendering. While the vast majority of the topics will focus on using data captured by RGB cameras, we will also discuss other means of capturing data using advanced sensors such as IMUs and event cameras. Finally, we will discuss quantum visual computing and the impact it can bring to the field.
Computer Vision strives to develop algorithms for understanding, interpreting and reconstructing information about real-world scenes from image and video data. Computer Graphics focuses on image synthesis: algorithms to build and edit static and dynamic virtual worlds and to display them in photorealistic or stylized ways. Machine Learning is concerned with studying and developing algorithms which use statistical models to solve problems by analyzing and drawing inference from data. In recent years, these fields have converged more and more. Both Computer Vision and Computer Graphics create and exploit models describing the visual appearance of objects and scenes, while the most successful models heavily utilize ideas from Machine Learning. In this seminar series, we will cover advanced research topics that cross the boundaries between the fields of Computer Vision, Computer Graphics, and Machine Learning. This seminar will cover research papers from the following topics: