Auto Camera Calibration

Question

From Wikipedia, camera cutting is the process of finding the true parameters of a camera that create a given photo or video. Camera cutting is also known as geometric camera calibration.

I am currently using the Camera Calibration Toolbox for Matlab to calibrate the camera. The instrument returns calibration parameters such as focal length, principle point, skew and distortion. However, the problem with this method is that it takes an extra step to calibrate the camera using a special calibration object such as a chessboard. In addition, it works for only one camera focus.

How can I get calibration parameters without manual calibration? For example, how does Microsoft Photosynth calibrate the camera on its images?

Answer 1

You are looking for a study called self-calibration, or auto-calibration . There are several articles (free), and I recommend starting with this tutorial .

Answer 2

Photosynth has the advantage of having multiple images of the same scene and can track points of interest through them. This is probably the main method they use to determine the locations where photos were taken, as well as viewing angles and focal lengths. Although you probably only get relative results for other representations, most of which are probably grouped in the same plane, then you simply declare them as grounded.

By the way: the researchers who built it made publications about it that are available online ¹ : Photo Tourism , Modeling the world from online photo collections , Finding paths through world photos .

---

¹ If you have an ACM subscription, but usually you should have / uni / whatever at startup.

Answer 3

Perhaps the camera manufacturer may provide you with data. I don't know anything about Photosynth, but any “calibration” performed without any object for calibration or the known properties of lenses and sensors, and thus should be based on suspicious previous beliefs, isn't it?

Edit: I see from other comments that Photosynth stitches photos together. Thus, previous beliefs include the knowledge that several photographs are images of different aspects of the same scene. Its task is not so much in calibrating the camera as in matching the images themselves.

Answer 4

You also do not need to do anything, like the level of sub-pixel correction that you get from the chessboard.
In the best case, he should just rotate and move the overlapping images, even with bad images he needs to find several edges to take out the converging verticals.