​

Could the wireframe data be distilled even further? Each wireframe image was distilled into a set of 19 joint angles. The transformation process now involved two phases: conversion of raw images into wire frames, and wire frames into joint angles. Taking three data points (each normalized with a reference to the wrist or “point 0”), the angle of individual fingers relative to each other was calculated. The use of joint angles added an additional dimension, the classifications would now be scale-invariant i.e., independent of the length of the fingers.

​

Because the input space was now so compact, using a simpler model fitting algorithm (Random Forests) was now feasible. As the name suggests Random Forests includes a forest of decision trees. Incidentally, the decision thresholds within the individual trees are not random but are carefully chosen to maximize the information gained by the separation boundaries. One advantage of the Random Forests model is that it is simpler, and because it is based on decision trees, the decision thresholds are well-suited for scrutiny and interpretability.

​

The Random Forests model was configured to have 100 decision trees and had an accuracy of 97%.

Image credit: https://www.britannica.com/science/hand-anatomy