Neural Rendering

Recent advances in neural rendering have made it possible to generate novel photo-realistic views of real-world 3D scenes just from a set of regular images. The most successful approaches combine ideas from computer graphics, machine learning, and optimization. Specifically, neural rendering methods such as NeRF or DeepFusion combine conventional volumetric rendering methods with a coordinate-based neural networks trained to predict continuous density and radiance estimates. These learned scene representations also have broad application across a wide set of domains, including virtual reality, robotics, health, and computer vision.