Free online courseComputer Vision: From Image Formation to Neural Radiance Fields

Course Description

Turn images into measurable geometry, interpretable models, and intelligent predictions. This free online course guides you through computer vision from first principles of image formation to modern neural methods such as Neural Radiance Fields (NeRFs). Instead of treating vision as a black box, you will learn why it is fundamentally an inverse problem: the world causes images, and vision must reason backward under ambiguity, noise, and missing information. That mindset connects classic geometry with today’s learning-based systems and helps you build intuition that transfers across projects.

You will develop a strong foundation in how cameras form images, how sensing choices affect what data you can trust, and how geometric transformations relate views of the same scene. From there, you’ll progress to 3D understanding: matching features across images, reasoning with epipolar constraints, and scaling reconstruction with optimization techniques used in practice. You’ll also see how stereo depth is computed in rectified settings, why local matching often fails, and how spatial regularization improves consistency when estimating dense disparity.

A key theme is structured prediction. You’ll explore probabilistic graphical models, Markov random fields, and factor graphs as tools for expressing dependencies in vision problems like denoising, stereo, and optical flow. That perspective makes it clearer how inference works, what the partition function is doing, and why message passing can be efficient when the structure permits it. You’ll then connect these ideas to learning: conditional random fields, parameter estimation, and modern deep structured approaches that unroll inference into trainable networks.

Beyond geometry and graphs, the course bridges into shape-from-X methods and multi-view fusion, helping you understand how light, reflectance assumptions, and sensor cues influence 3D recovery. You’ll then step into implicit neural representations and differentiable volumetric rendering, where models learn 3D structure and appearance by backpropagating through the rendering process from RGB images. NeRFs and generative radiance field ideas show how view synthesis and reconstruction converge in contemporary machine learning.

To round out your toolkit, you’ll work through recognition tasks such as classification, semantic segmentation, and object detection, and study self-supervised and contrastive learning for reducing reliance on labels. By the end, you’ll be prepared to read research with confidence, prototype robust pipelines, and speak the language of both classical vision and modern AI—useful for careers in machine learning engineering, robotics, AR/VR, and 3D content.