Implement DETECT_PARTIALLY

[qbonnard]

(name negociable)

The idea is to combine detection and tracking, where detection whould happen only on a subregion of the input image. The point is to have a faster detection/tracking, which would happen in a more constant time than DETECT_PERIODICALLY.

Simply put, DETECT_PARTIALLY would start the detection on the top left quadrant, and tracking on the full input image. On the next frame, the same would happen with the top right quadrant, then the bottom left, bottom right.

In practice, the subregion would have to slide more smoothly to avoid misdetection of tags on the borders of the quadrants. We also need to take the full input image to handle big tags. Ideally, we could set the size of the subregion as a parameter.