Our Client:

Prospective insurance company

Situation:

Insurance claim reports require the evaluation of damage done to cars to assess how much compensation should be rewarded. Manual evaluation can be time-consuming, and insurance companies are looking for ways to streamline this process. DSL have developed an instance segmentation model that is able to detect and classify any such damages.

Task:

The task was to design and implement an instance segmentation model, that was able to precisely detect and classify different car defects specified to the model. Here, the model’s precision is very important as it aims to map out exactly where the damage is, as opposed to a simple box approach to identify the fault. To be able to achieve high quality segmentation properties, as well as learn to distinguish between different faults, the model would require to have a large amount of instances of each class to learn well.

Dataset quality is a crucial component of any machine learning pipeline and through the use of publicly available data, we were able to obtain a dataset with 10k images. The use of state-of-the-art architecture aids in achieving our aims of a high precision model, and thus we opted for the Mask2Former model. This pre-trained model was further trained and fine-tuned to our custom dataset, to achieve a greater performance in our scenario.

Using state-of-the-art architecture will aid in achieving the high precision we sought after, and thus we opted for the Mask2Former architecture. This was further fine-tuned to the specific use-case using our custom dataset.

Result:

The results were good and as a whole the model performed as we intended to. Some edge cases to consider are that the model struggled to correctly segment things such as small dents and scratches. Factors such as camera angle and quality attribute significantly to this as it can make it that much more difficult to identify, and a poor angle choice and further harm the decision process. Scratches are not as big of objects as a flat tyre for example, and this are more subtle and harder for the model to segment.

Expanding on the edge cases, with publicly available data being easy to access also caused disadvantages. One challenge we faced was inconsistencies within the image dataset, whether that is environmental factors or quality of image captured. This issue can be mitigated by making small incremental improvements to the model, and working directly with clients and understanding their work environment.