EuroSciPy 2024

A Comparative Study of Open Source Computer Vision Models for Application on Small Data: The Case of CFRP Tape Laying
08-29, 16:00–16:30 (Europe/Berlin), Room 7

The world of open source computer vision has never been so exciting - and so challenging. With so many options available to you, what's the best way to solve your real world problem? The questions are always the same: Do I have enough data? Which model should I choose? How can I fine-tune and optimize the hyperparameters?

In collaboration with the German Aerospace Center, we investigated these questions to develop a model for quality assurance of CFRP tape laying, with only a small real data set fresh from production. We are very pleased to present a machine learning setup that can empirically answer these questions. Not only for us, but also for you - our setup can easily be transferred to your application!

Dive with us into the world of Open Source machine learning tools that are perfectly tailored for your next project. Discover the seamless integration of Hugging Face Model Hub, DvC and Ray Tune. You'll also gain unique insights into the fascinating world of CFRP tape laying, specifically how well different architectures of open source models perform on our small dataset.

If you want to level up your MLOps game and gain practical knowledge of the latest computer vision models and practices, this talk is a must for you. Don't miss the opportunity, and look forward to your next computer vision projects!


The world of open-source computer vision has never been so exciting—and so challenging. With so many options available, what's the best way to solve your real-world problem? The questions are always the same: Do I have enough data? Which model should I choose? How can I fine-tune and optimize the hyperparameters?

In collaboration with the German Aerospace Center, we investigated these questions to develop a model for quality assurance of CFRP tape laying, using only a small real dataset fresh from production. We are very pleased to present a machine learning setup that can empirically answer these questions. Not only for us, but also for you—our setup can easily be transferred to your application!

This talk provides you with a blueprint for your own projects, focusing on a setup that allows you to improve your models in a controlled manner and compare results effectively:

  • We begin by examining the problem through our specific use case of CFRP tape laying and breaking it down into generic solution steps.

  • These solution steps are translated into a machine learning pipeline using DVC (Data Version Control). This approach saves computation time on steps where neither the data nor the source code has changed and helps to keep track of your progress and performance over time using Git.

  • We will explore various current model architectures available in the Hugging Face Model Hub and demonstrate how you can fine-tune them on your data using Python packages such as transformers and ray. On the topic of hyperparameter search, we will discuss the available algorithms and the most promising parameters.

  • Finally, we will review our results, specifically how well different architectures of open-source models perform on our small dataset. We will explore the question of how different model architectures compare and whether the largest model always gives the best results.

If you want to level up your MLOps game and gain practical knowledge of the latest computer vision models and practices, this talk is a must for you. Don't miss the opportunity, and look forward to your next computer vision projects!


Abstract as a tweet

The world of open source computer vision has never been so exciting - and so challenging. Level up your MLOps game and gain practical knowledge of the latest computer vision models and practices, don't miss the opportunity!

Category [Machine and Deep Learning]

ML Applications (e.g. NLP, CV)

Expected audience expertise: Domain

some

Expected audience expertise: Python

some

Public link to supporting material

https://github.com/pd-t/crfp-transfer-learning

Meet Thomas, a passionate advocate for science, particularly in the realm of applied mathematics. Following his doctoral studies, he embarked on a journey into the world of embedded programming, where his affinity for DevOps took root. His enduring passion for crunching numbers ultimately led him to the fascinating field of artificial intelligence, where he's now an acknowledged MLOps expert, seamlessly integrating machine learning into operations.

Thomas has an impressive track record as a leader, having overseen two publicly funded open-source research programs in the field of AI, in collaboration with the German Aerospace Center. Today, he is at the forefront of AI-driven cybersecurity research at Smart Cyber Security GmbH and working on his low-budget bark beetle detection drone project – a testament to his enduring fascination with embedded systems.

Tim is a ai engineer at WOGRA AG, which is based in Germany.
More information about the speaker will follow.
Stay tuned!