This page lists what this solution needs: the libraries, the machine, the data and any licence.
1. What it needs#
The list below follows the method from the earlier pages, one need at a time.
No labelled data. Nothing in the method is fitted. So it never reads the examiner's training half of the arrangements.
No training and no weights file. There is nothing to train, and nothing to keep up to date with the cell. So there is also no licence to follow.
No graphics processor. The work is simple arithmetic on a picture of 320 by 240 pixels. It compares heights, back-projects the kept pixels, drops one column of numbers, grows a grid of 5 mm squares, and does one direct solve per group. An ordinary processor does all of it.
Two libraries, both already in the cell. NumPy does the arithmetic on the depth readings. OpenCV grows and joins the marked squares into groups. It also finds the outline of a patch of dots before a circle is fitted to it.
Three facts from the problem, not from the sensor:
- the table's height, which the cell knows because the table is bolted to the arm's frame;
- the smallest distance between two glass centres, which sets the upper limit on the grouping distance;
- the range of widths this kind of glass can have, which the width check uses.
Depth readings. This is the one need that is not free. How it compares explains why.