Overview

Summary:
  • NeRF models the volumetric distribution of light rays in a scene, allowing it to generate novel views with high visual fidelity. It achieves this by training a neural network to predict the radiance at any given 3D point and rendering views by ray marching through the learned radiance field.
  • Poses and camera parameters: Obtained by using Metashape software.
  • Images: Captured using an iPhone camera keeping the focus and exposure constant.
  • Drawbacks: Requires large amount of data, slow to train, sensitive to image and pose noise.
Poses and camera parameters: Obtained by using Metashape software.
Images: Captured using an iPhone camera keeping the focus and exposure constant.
Drawbacks: Requires large amount of data, slow to train, sensitive to image and pose noise.
Paper: NeRF
Results:: The below results are for a Vanilla NeRF model trained on the images captured using iphone.

Full RGB Rendering

Depth rendering

Insights

  • The method transforms the 3D point to Normalized Device Co-ordinates(NDC).
  • Refer the supplementary material.
  • For perspective projection matrix, I referred this blog. overview