Created a web-app to take input as images from an user and to classify those images as damaged(multiclass i.e. minor,moderate and severe). Given less data which was scarped from google images using selenium(script uploaded in the code directory), the model performed well(comparable to an avg human). All the models related figure i.e. accuracy,precision,recall is reported in the pdf. We trained a pipeline of convolution neural networks(CNN) model using transfer learning on pretrained vgg16 model(i.e. removing it's fully top-connected layer) using keras and tensorflow.
Web app - [Car Damage Evaluation]() Updating later...
Get the data and pretrained models and more from this drive link
Get the updated model in json format from here