Lecture 4

Question 1

What is the pose of a camera? What method can be used to compute this from 3D point locations, 2D image locations, and camera calibration information?

Answer 1

The rotation and translation. Perspective-n-point pose can be used to compute these.

Question 2

What is perspective-n-point pose? What do we need?

Answer 2

In the minimal case we need 3 points which are part of a 2D image like a checkerboard, if we have more images than this we can simplify the algorithm, use least-squares fit, and remove bad matches.

Question 3

How do we refine the pose given by PnP?

Answer 3

Generally R will not be quite a rotation, we can compute the SVD to get the U and V components, which are the rotation, we can set the sigma component to 1 to isolate just the rotation.

Question 4

What is RANSAC?

Answer 4

Robust pose estimation can be used when we have more than 3 matches, it works by

1. sampling 3 matches at random
2. Estimating the camera pose using Perspective 3 point pose.
3. for each pose estimated count how many points agree this is the consensus set.
4. repeat steps 1-3 many types
5. keep largest consensus set.
6. make a least squares fit to it.