Car damage detector

This application is made to detect which part of car is damaged

To quickly view final results open Final output and scroll to the bottom

Dataset:

• Training set: 62 images
• Validation set: 8 images
• Test set: 8 images

Approach:

The nature of annotations in the dataset and its small size make segmentation a good option.
FPN model available in pytorch is used. Initialised with ResNet weights for feature extraction
Two segmentation models are trained:

1. To give which parts of car body are in image
   This model is trained on the 5 parts tagged in the annotations
2. To give which area is damaged

IOU over these two results tells us which body part was damaged.

On test set of 8 images, 7 are classified correctly.

Further:

Multi task learning could be attempted or models could be combined to create single model.
But couldnt try due to time constraints