Personal notes on Structure from Motion (SfM) - Technique to recover 3d structure from 2d images
Step 1: Extract features from images
First, we extract sift features from the set of images. These features are used to construct the scene graph. Iteratively, we start with a pair of images with the most amount of common inliers, and keep adding new images to the scene graph.
Step 2: Triangulation
Using all pairs of images, we triangulate the 3d points to construct the sparse cloud of 3d points.
Step 3: Bundle adjusment
Finally, we conduct bundle adjusment of the 3d points by calculating the reprojection error of the 3d points, when mapped back to the 3d images.