Perspective and Homography

*** Work in progress ***

Challenges from the perspective (homography) problem involve:

Estimating orientation with respect to the floor plane is not trivial. Typically visual markers indicating 3 points with known relative positions on a plane would suffice (or 4 if counting scales). An example is the 3 corner markers in a QR code. But there is no good place to embed such markers on the Code Map because it’s a continuous code that already uses every bit of space. Embedding higher details is also difficult because it’s already designed to operate at the limits of camera resolution. Without adding visual markers, it is still possible to estimate the perspective by comparing how much the grid lines deviate from perpendicular squares, but such a procedure would be a resource intensive statistical approach. Potentially a simple hardware solution could be adding an IMU, but that would require dedicated hardware, and generally increase costs. Even if the correct perspective/orientation/homography was provided by an oracle. To undo the perspective transform requires the source image to be of sufficient resolution in the first place.

Scale search

• relatively cheap in 1D, even for exhaustive search.