[CVPR 2022] NeRF - Neural Radiance Fields for View Synthesis

NeRF is a methodology that innovatively solves the View Synthesis task of generating unseen viewpoints using images captured from multiple angles. It achieves performance superior to existing methods through a simple MLP structure that takes a 5D representation (x, y, z, θ, ϕ) as input and produces the RGB values and Volume Density (density/opacity) at that location.

NeRF-related research presented at CVPR 2022 can be broadly classified into four categories.

• Single Image-based NeRF: Technologies that generate new views using just a single image and an estimated camera pose, rather than multiple views (LoLNeRF, Pix2NeRF, etc.), have emerged.
• Overcoming the Limitations of NeRF: Studies have been published that address slow optimization speed and noise issues by combining Point Cloud, improving background blur through Regularization, and handling complex reflections.
• Scalability and Geometric Consistency: Technologies for reconstructing large-scale scenes, such as Block-NeRF, which trains city-scale scenes by dividing them into block units, have drawn attention.
• Handling Non-rigid Data: Research continues on handling non-rigid changes such as human poses, and on implementing complex reflections of glass and mirrors.