Building AI-based solutions for computer vision problems arising in the second-hand car market

Ninety percent of all used car purchases start with a search on the internet. Of course, anyone looking for a used car on the internet would like to know a lot about the car. Twinner, a German company headquartered in Halle launched the digital car scanner Twinner Space in 2018 and their system is already in use in several companies and car dealerships in Germany and China. The vehicle is driven inside a scanner and within a few minutes, it generates a detailed 360-degree image of a vehicle. The scan can assess damage to the bodywork, detect repaint, and it can even measure the tyre treads without costly and time-consuming human intervention. The digital scans, the 3D geometry models and the detailed selected views produced by object detection algorithms increase customer trust and can serve as evidence in the event of insurance claims.

In this talk I would like to explain how AI-based object detection and semantic segmentation algorithms can help us to solve problems arising in car digitization. Convolutional networks detect various objects of interest on a car by providing bounding boxes that enclose the object. They can be used for automatic detection of car body parts, that is, they can make pixel-level decisions and locate parts of the car that is a useful feature for automatic damage detection. I explain how the data science team train and use such models for inference in various deep learning frameworks like Tensorflow or ChainerCV and I also give an outline how all necessary components and data pipelines are implemented using dockerized Python microservices deployed in a Kubernetes cluster.

Dr. Kurics Tamás
Data Scientist, Twinner

Tamás has graduated with a master degree in applied mathematics at Eotvos Lorand University in Budapest. He obtained his PhD in mathematics in 2011 and did his postdoc in computational biology at ETH Zurich in 2012 and 2014. He left the academic world for the industry in 2015 and joined Balabit, an IT security company where he worked as data scientist in a team responsible for developing anomaly detection and biometric algorithms. He joined Twinner in 2019 where he works in the automotive industry, responsible for prototyping and implementing deep learning models and data pipelines with his colleagues in Budapest and Berlin.

He is a frequent problem solver at online competitive coding platforms. His main interests are the applications of mathematics and computer science in life sciences, functional programming languages, particularly Scala, object detection and machine learning in computer vision.