The cost of your car insurance is typically calculated based on the miles you drive. But that doesn’t always have to be in your best interest, as the number brackets given are generally just a rough estimate usually costing you lots of money. One car insurance company approached us to see, if we could develop a pay-per-mile system, which would heavily rely on reading vast amounts of dashboard images (odometer readings) to capture the actual car mileage, while incorporating security measures to identify any potential manipulation with the picture (fraud prevention).
To minimize (if not eliminate entirely) any human performed odometer readings of dashboard images taken and uploaded from smartphones, we developed a cognitive system based on DNNs, which:
Two weeks into the project, we had developed a viable MVP and only a few weeks later we finalized Odocap with a reading accuracy of 99.9%. This was pretty exciting given the fact that the images were of different quality, taken in different light conditions and from all sorts of angles, which made the process even more complex and interesting. The car make and model accuracy was at 98% (system is trained on 25 different car makes and models - mostly European).
In case of interest, it takes only several days of deep learning to add additional car makes and models.
After 10 weeks, we finished Odocap, which is cutting human work-time spent on image evaluation by a staggering 90%. Furthermore, our system is 100% consistent and error free.
The Odocap can be deployed on premise or on cloud with a mobile version, which is also available on request. If you are interested in using Odocap, contact us.
Our system still experiences some difficulties in recognizing a Bugatti Veyron make and model :-), but that is solely due to not having enough data. Should you have several tens of Veyron’s dashboard images, please share them with us so we can improve Odocap’s performance!
ODO is a deep learning program based on OCR of car dashboard images taken by smartphones to monitor car mileage.
TensorFlow, Python, C++, CxFlow