Using our dented “mini-hood” and our Quidient Reality™ prototype, we previously demonstrated the ability to create 3D dent profiles from a single image that measured within 28-micron RMSE in comparison to a highly accurate Zeiss Comet structured light system. We also provided a method for using this measured profile to determine a reference size (SR) suggested to be 6 times the standard deviation of the profile, hence six-sigma. We now discuss how we used machine learning (ML) to train Quidient Reality™ to automatically measure dent size patterned after industry experts. Our first step was to invite two hail dent assessment experts to our lab at Quidient’s Technology Center to visually measure dent size (SV) for all 57 dents on our mini-hood, multiple times from different orientations. Their SV measurements where then processed along with Quidient’s SR measurements using ML to determine a learning coefficient “k.” We then show that by multiplying our automatically determined reference size SR by the learning coefficient k, Quidient Reality™ is able to repeatably measure dent size SA that is on average nearly identical to an expert’s visual measurements SV. And finally, we showed that our resulting automatically determined dent size SA is convertible a “coin size” following accepted industry practices.
Our mini-hood comprises 57 total dents of various sizes distributed over the four colored quadrants. We conducted independent measurement sessions with each of two industry experts. In each session we first asked the expert to mark the size of each dent with a grease pencil defining what in their opinion was the boundary width of the “claimable dent.” Our team then measured each of the expert’s boundaries associated with the respective dent to obtain the visual dent size SV.
After cleaning off the mini-hood we asked each expert to repeat the process this time from a 90-degree rotated perspective. We repeated these steps to create second sets of measurements at each rotation and compiled the expert measurements for each of the various mini-hood dents as shown in the table on the left.
We then used ML to determine a best fit between the expert’s visual measurements SV and the Quidient Reality™ reference measurements SR. Although several ML algorithms could be used, we chose the linear least squares method to arrive at the learning coefficient “k,” where in our case k = 1.06. Next, we calculated an automatically determined size SA = k x SR. As shown in the Dent Size table below, after scaling SR by k, the resulting SA on average is identical to the industry expert’s SV.
(We note that our reference size SR was automatically determined by first sampling deviations along our measured 3D profile (the dent) as compared to a Normalized (undented) panel surface, mapping these deviations into a 2D bell curve diametric distribution. Using the resulting distribution, we set SR = six-sigma, thereby statistically representing 97.7% of the profile deviations from the Normal.)
The table on the right compares the expert’s average visual dent size’s SV with our automatically determined sizes SA. Across the entire 57 dents, our average SA = k x SR result substantially equaled the expert’s SV, where our largest deviation was roughly 30%.
And finally, we asked our experts to classify each of the mini-hood dents according to the industry’s traditional “coin system” where dents are sized as “dimes, nickels, quarters or half-dollars.” We then used the industry standard Dent Size Chart shown above threshold our dent sizes SA measured in millimeters into the conventional coin system. Our resulting coin sizes matched those of the experts.
We have demonstrated a useful method for calibrating Quidient Reality™ to measure hail dent sizes “trained” like an industry expert. Our results have shown that Quidient Reality™ will allow any conscientious person such as the vehicle owner to measure dent sizes on par with industry practices. We further anticipate that the P&C industry will eventually establish a process to determine an industry-wide learning coefficient k, where Quidient Reality™ is then trained across all insurers to provide standardized dent size SA measurements. Our vision includes a Hail Dent Assessment Gateway connecting our platform with all industry stakeholders opening up an era of owner-initiated and first-assessed claims processing. We believe that in the near future a vehicle owner using the Quidient Reality™ Engine embedded in their cell phone will measure their own vehicle damage including hail dents and other non-structural damage.