The Problem: Given a series of views of a static scene, all taken from the same camera with unvarying internal parameters, calculate the internal camera calibration from the views alone, without any extra information about the scene (such as measurements in the scene).
Example: The six computer-generated views below were used as input to my algorithm. They show three squares that meet at a corner at right angles to each other (like the corner of a room). Each square is covered with a texture to give an automatic point-tracking algorithm something to track.
Points were automatically tracked between the images using an off-the-shelf point-tracking program. A fundamental matrix between each pair of views was found using a RANSAC approach, and these fundamental matrices were used to find the internal camera calibration.
Once the internal calibration is known, metric scene reconstruction is possible (metric scene reconstruction is identical to the original scene up to an overall scale factor). Several view of the reconstructed object are shown below:
|
|
|
|