Figure 2: Keypoint tracking and feature extraction.

a, 17 roughly equidistant points were hand-labeled along the length of the arm. Using these keypoint positions, various metrics were computed for subsequent analysis including the overall angle made between the (stationary) base and the tip (θ) and the angular and keypoint velocities. The distal-most keypoint (x) and its velocity (v) were found to be significant in distinguishing motion types. b, To quantify motion across time, three intervals post-stimulation were considered: the first (t0) and second (t1) seconds, where most motion occurred, and 2 or more seconds (t2) until any observed motion ceased. c, The histogram of one example metric, the maximum angular velocity in t0, is plotted (using a kernel density estimator) for each of the human-labeled movement categories. As expected, the ‘No movement’ videos have very low or zero angular velocity, whereas ‘Movement with arm curl’ videos tend to have higher maximum angular velocity.