Currently it's not quite at the level of graphics tablets, but it's not too far off, and I think there's quite a bit of potential to improve it using similar te

Currently it's not quite at the level of graphics tablets, but it's not too far off, and I think there's quite a bit of potential to improve it using similar techniques [1]. In terms of absolute accuracy, I measured an average error of 0.89mm (for the position of the tip) across the entire area of an A4 page with the camera in one place. In practice you have more precision than that though, because most of the errors are constant biases (not random noise). For example, here's one of the tests [2] I did for the thesis, which compares the recorded stroke to what I actually wrote (scanned from carbon paper). After aligning the two captures (as a global 2D position offset, everything else is retained), the average distance from the recorded stroke to the scan was 0.158mm. [1] This paper (which I linked in another comment) uses some more advanced techniques for pose estimation which could definitely be applied here (but it's closed-source, and I didn't have time to re-implement it from scratch): http://media.ee.ntu.edu.tw/research/DodecaPen/ http://media.ee.ntu.edu.tw/research/DodecaPen/ [2] https://github.com/Jcparkyn/dpoint/files/13329235/main-sketch-plot.pdf https://github.com/Jcparkyn/dpoint/files/13329235/main-sketc...