During my summer stint as a part time worker at a local Car rental company, I noticed that my supervisors were struggling to get enough workers to identify the damage on the arriving cars, mainly because around 150 cars arrived per hour and getting to notice every single dent on them was cumbersome. Being an AI pioneer myself, I realized the potential impact of AI for not only my supervisor, but also to hundreds of Car rental, car wash and insurance companies out there which could make enormous efficiency gains by adopting a software which automatically detects and marks the dents for them. It was then that I decided to get together a team and start working on this project!

What it does

This project introduces an innovative approach to detect defects cataloged as scratches, dings and dents on vehicle body surfaces, which is currently one of the most important issues facing quality control in the automotive industry. By taking inputs as .jpg or .jpeg images, the application returns a new set of images with the dent/damage regions clearly marked out, along with the percentage of confidence with which the application makes the claims.

How we built it

We used the Mask RCNN implementation of Tensorflow 2.x and trained it using 1500+ images of damaged vehicles. We used VGG image annotator for labelling the damaged portions. We then trained the model for 4 seed epochs and 8 main epochs for a total of about 33 minutes, and exported it into .h5 format so that it can be easily integrated with front end applications.

Challenges we ran into

1) Mask rcnn implementation was not available for the latest version of tensorflow (2.x). So we had to work around and get an unofficial implementation of the model.

2) We found it difficult to train 1500+ images as it took a lot of time.

3) Most of the Tensorflow tutorials on the web were not updated to reflect the latest changes in it. So most of the tutorials did not work satisfactorily. Hence a lot of trial and error methods were required to connect and use the appropriate versions which would blend together and work perfectly.

Accomplishments that we're proud of

Automated detection of car exterior damages and subsequent quantification(damage severity) of those would help used car dealers(Marketplace) to price cars accurately and fast by eliminating the manual process of damage assessment. The concept is equally beneficial for property and casualty(P&C) insurers, in terms of faster claim settlement and hence greater customer satisfaction.

What we learned

By building this project we learnt how to implement Mask RCNN and use VGG image annotator for our projects. But more importantly, we learnt the efficiency gains from teamwork, where dividing the work according to specializations proved to be extremely wise as we finished the project in 24 hours what would normally take many weeks for individual contributors.

What's next for Vehicle Damage Detector.

We plan to improvise on this software by making websites, mobile apps, and many more for the public's usage. We also plan to pitch this software to various Car rental companies, Insurance companies, Car wash companies and make their operations faster and efficient using this AI.

Share this project: