In the past year, Fit3D has been collecting more and more body measurement data from Fit3D scanners all around the world to continually refine and improve its algorithms. The body fat prediction model we developed earlier has been steadily improving with the addition of more training data based on a combination of Fit3D scans and DEXA scans.
The graph below shows the Root Mean Squared Error (RMSE) for body fat, which is a measure of the average error in body fat prediction. The error bars on each point show the standard error in RMSE for each training set size (N), based on performance over 10 bootstrap samples, with each run tested on a common hold-out test set.
Figure 1: Change in Fit3D body fat percentage prediction model’s performance by training set size. Our original model used a training set of size 1041.
We’ve been encouraged to see a steady decrease in error as Fit3D collects a more diverse dataset of human body measurements representing people of all shapes and sizes.