In this project, we sought to train a computer vision model to detect the location and severity of car damages in order to give an estimate of the cost of repairs, which could help insurance firms process claims faster and detect fraudulent claims. For the model training, we gathered around 140 images of damaged cars with either scratches or dents. To train the damage detection model, we identified scratches and dents in images and labeled them with VGG Image Annotator.
For the training of the model, we fine-tuned a Mask R-CNN Model that was trained on the MS COCO object detection problem as a starting point to finetune the model with our dataset of damaged cars. With a very small dataset at around 140 images, we decided to augment our images to add more variety to the dataset. This helps the model in capturing some of the patterns under different angles, different lighting conditions and different image resolutions.
For the location detection, we trained another Yolov5 model on 3 classes, namely front, back and side. We did not train location detection using Mask R-CNN as we only needed to know whether the damages are front, side or back and hence instance segmentation was not really necessary.
Using Flask and REST API, we integrated the MaskRCNN and YOLOv5 models into a web application that enables end users to upload car images and receive model inference results.
Article:https://medium.com/@nusfintech.ml/car-claims-insurance-technology-insurtech-4ffbe2c948c5Members:Daniel, Amateus, Chester, Nielsen, Zhili