Crash detection with Class Activation Maps

1. Goal

The goal of this repo is to classify car crash images into crash and non crash. Additionally, it should be shown where the crash is found on the image. This is done via Class Activation Maps (CAMs) without gradients. This allows for localisation of the crash without pure object detection algorithms.

2. Structure

2.1 Training of the model

The model is trained using PyTorch and a config file. With the config file, new models can be built very easily using a setup like this:

resnet:
    epochs: 100
    
    batch_size: 16
    learning_rate: 0.001
    train_backbone_params: True
    train_test_split_ratio: 0.8
    activation_function: 'tanh'
    classifier_layer:
        - classifier_layer_1:
            - 512
            - 256
            - 0.2 # Dropout rate
        - classifier_layer_2:
            - 256
            - 6
            - 0

• The model can be trained in the Azure Cloud if the correct settings are made and the Azure connection is established.
• The training process is monitored via MLFlow.
• If possible GPUs will be used.
• The final model class can be used to export the trained model. It contains a method that computes the CAM of the feature map of the last convolutional layer.
• The model has an validation accuracy of 100%.

2.2 Deplyoing the model

The fully trained model is deployed on some local endpoint. This will be improved in the future.

2.3 Example output

After training the model predicts wether there is a crash or not. The activation map shows where the crash was found on the image.

2.4 Deployment

Currently the interactive session does not work, because it costs to much to host the backend in the cloud. Nevertheless I have an docker image of the backend which you can find here: https://hub.docker.com/r/coluding/cam-backend.