Skip to content

3D Reconstruction Pipeline (Structure from Motion)

image-20240506140841559

The following is the basic workflow to recover 3D data:

image-20240506141013696

The algorithm below assumes that the structure is static.

Structure from Motion (SfM) vs Simultaneous Localisation And Mapping (SLAM)

image-20240506150104341

Get Point Correspondences

When talking about still images, one usually talks about structure from motion. On the other hand, when taking about finding correspondences in a video, this is called SLAM. However, both use the same basic algorithm.

Finding Camera Position

To get the camera position, the foundation matrix \(F\) needs to be recovered. See Intro/Stereo to see the details.

image-20240506141907988

The steps are \(F \to E \to R,T\) to get the rotation matrix \(R\) and the translation matrix \(T\) of the camera.

Scene Structure: Triangulation

TODO

In praxis, usually there will not be a point that satifies both contraints, since there is noise. This is solved by using a direct linear transform, which performs a least square problem.

Sequential Structure from Motion

image-20240506142550474