INITIAL SITUATION

European Emission Laws requires advertisers to label the car model in any photo shared on social media.

Our client would like a way to automate the labelling of car models when they are copied from personal accounts to the company account for advertising purposes.

SOLUTION

A deep convolutional neural network was trained to identify car types.
It can identify roughly 200 different types of cars with near-human level accuracy (85%).

BUSINESS VALUE

The client is able to automate the process of labelling cars, which frees up the company’s employees’ time for more valuable tasks.

Cars can be classified much faster than with a human labeler (a few milliseconds, compared to 1+ minutes for humans).

FRAMEWORK & TOOLS

The project was developed using Keras with a Tensorflow backend.

It was run as notebooks on Databricks, managed via Azure Machine Learning, and exposed for business process integration as a REST web service.

An inference program hosted on AWS Sagemaker could be queried with an image url, and would return the label and associated probability.