Streamlining Car Damage Detection

Image annotation

The client was working on an AI solution that could not only detect damage caused to cars but also provided the clients with additional services such as submitting a claim, appraising the damage, predicting the type of repair required, and a lot of other useful information. However, to ensure the accuracy of the AI system, they needed to annotate a dataset consisting of 36,000 images. The quality of labeling and geometry precision of 2D Bounding Boxes was crucial since any deviation could change the nature of the damage. 

We needed to detect all of the various types of car damage after accidents (scratches, crashes, broken parts, etc.) and classify them by the required taxonomy. It was important to detect all damages and allocate each of them by 2D bounding boxes. Some images contained many objects for annotation of different types, and annotators needed to be extra careful in object recognition and classifying.

Since Mindy Support has a solid background in projects for the Automotive and Insurance industry as well as experience working with the needed subject matter, the client approached us for assistance in providing the needed data annotation services. Our portfolio and experience in similar projects were important for the client, as was our ability to provide personalized recommendations about the necessary data annotation tools necessary to complete the project. This freed the client from having to delve into minute details of having to choose a data annotation tool, and they could just focus on providing input data and getting output data in the required format.

Mindy Support took full ownership of the process development of the data annotation project, selecting an annotation platform and process management. The client requested to play a minimum role in all phases of the data annotation process. This included things like team recruitment, training, tools selection and many other processes. In order to accommodate the client’s request, we only gathered from them instructions and the data required to complete the project. We then took over the entire process of training the team, as well as quality control and generating the output of the desired format. It was important for the client to get a quality score higher than 98%, and several stages of annotation validation and quality control were carried out on the project. 

The client was very satisfied that all of the work was done on time and with the necessary accuracy level, both in geometry and labeling correctness. Since the first stage of the project was successful, the client decided to allocate additional data annotation work to us required to complete the project. We are currently waiting for an additional couple of datasets with a much bigger diversification of object classes.

• 36,000 images annotated
• 98% quality score achieved with minimal calibration on the annotation approach by the client.
• Project duration: 2.5 years