Structure from motion exploration

With just two frames

feature matching

point cloud reconstruction is

With 13 frames it is

This is a very basic rudimentary SfM. For the same 13 input images Colmap reconstructs both sparse and dense in 2-3 minutes, which is incredible in quality and speed:

COLMAP classic

Sparse

Dense (Poisson)

Volumetric rendering

I ran the very same images and camera poses through NeRF and Gaussian Spatting.

NeRF

Gaussian Spatting