LG AI Research 237

NeRF (Neural Radiance Fields), which has emerged as a core technology in the computer vision field since ECCV 2020, implements View Synthesis technology that generates unseen viewpoints from images captured at multiple angles.

Unlike existing 3D rendering methods that directly stored color values in Voxels and thus suffered from low memory efficiency and quality, NeRF innovatively solved this by utilizing an MLP (Multi-Layer Perceptron) structure to compute the density and color of a specific Ray.

At CVPR 2022, various studies utilizing NeRF were presented, and the main trends are as follows.

• Single Image NeRF: A technology that synthesizes new viewpoints or estimates Depth using just a single image and a rough camera pose, without the need for multiple photos
• Overcoming NeRF's Limitations: Research that combines Point Cloud or incorporates traditional TSDF-based 3D reconstruction methods to address slow optimization speed and vulnerability to noise

These technical advances show that NeRF is establishing itself as a core model for practical 3D reconstruction and high-quality rendering, going beyond simple view synthesis.