This project aims to help estimate the damage done to a car after an accident. This API leverages state of the art deep learning models in order to assess the damage done to a car and its intensity from images of the damaged vehicule.
In this project I combined images from different datasets to come to a larger and more diverse dataset: Pletarion dataset: https://peltarion.com/knowledge-center/documentation/terms/dataset-licenses/car-damage Open Data Commons Attribution License dataset; Coco Dataset
The different datatsets were labelled differently, so I had to integrate a data labeling step, so I developed a small desktop GUI to speed up the labeling process:
During the training process I used a pre-trained xception architecture and adapted it to our particular use-case. Added to it a data augumentation layer and droupout layers for regularization (as well as L2 regularization).
The metrics used in the validation step are accuracy and f1-score:
Class | Accuracy | F1-Score |
---|---|---|
Bumper_minor | 90 | 88 |
Bumper_severe | 89 | 85 |
Door_minor | 99 | 98 |
Door_severe | 99 | 97 |
Body_minor | 79 | 81 |
Body_severe | 85 | 75 |
Glass_shatter | 94 | 81 |
Lamp | 85 | 75 |
Tire | 94 | 35 |
Mirror | 96 | 36 |
Detailed walkthrough the different steps and processes of the project:
https://colab.research.google.com/drive/1kGv-hCy6PUIVtjOVNMZNmDtQLvk0Gssd?usp=sharing