Machine Learning for Inverse Graphics: Reconstructing 3D Scenes

Traditional computer graphics turn 3D models into 2D images, but inverse graphics does the opposite—using machine learning to reconstruct the 3D world from flat images. Understanding this intersection of computer vision and graphics is key to modern AI applications in robotics, virtual reality, and spatial computing. Through this text-based course, you will grasp the core mathematical and conceptual foundations needed to build and train machine learning models that understand 3D geometry, lighting, and materials from 2D pixel data. You will learn to: Understand how cameras project the 3D world onto 2D planes using coordinate systems and camera models; Represent 3D shapes and scenes using voxels, meshes, point clouds, and modern implicit neural representations like Neural Radiance Fields (NeRFs); Apply deep learning techniques to reconstruct 3D geometry and textures from a single 2D image; Explore self-supervised learning methods to train inverse graphics models without massive labeled 3D datasets; Analyze geometric deep learning principles to ensure models generalize across different shapes and viewpoints. The course starts with essential 3D coordinate mathematics and camera projection geometry before moving into deep learning architectures for shape representation. You will progress through written explanations and structured code snippets that demonstrate how to implement these algorithms step-by-step. This course is designed for software developers, data scientists, and students new to 3D computer vision who want a solid conceptual and practical foundation without needing prior experience in advanced graphics programming. Start reading today to bridge the gap between 2D pixels and 3D understanding.