Stereo geometry

Question 1

Why would we want multiple views in images?

Answer 1

Some information is lost in a 2D image of a 3D space, mainly depth.

Question 2

What cues indicate depth?

Answer 2

Shading
Perspective effects
Motion
Texture

Question 3

What is stereo imaging?

Answer 3

Use two cameras at slightly different angles to determine depth (similar to human eyes). You use the disparity (change of pixel location from one image to the other) to triangulate and get the depth.

Question 4

What are the steps of Stereo geometry?

Answer 4

Calibration
Rectification
Stereo Correspondence
Triangulation

Question 5

What are some of the possible steps to calibrate a camera?

Answer 5

Use extrinsic parameters, then intrinsic coordinates
Use a scene with known geometry
Undistort images
Non-linear methods
Geometric calibration
Color calibration
Radiometric calibration

Question 6

Explain rectification.

Answer 6

You want to modify the two different images so that they are similar (in terms of angles, rotation etc). You can use epipolar geometry to do this.
Since a single 2d pixel can have many 3d locations (on a line), you can find the corresponding line in the second image.

Question 7

Explain Stereo correspondence?

Answer 7

The actual matching of images. You use a sliding window like with HOG, try to find where the epipolar line from one image matches best. You want to find the ground truth.

Question 8

How can Stereo correspondence be improved?

Answer 8

Assume that depth changes smoothly, and use energy minimization. You want pixels that match the other ones and move only slightly.