Thesis Work - Data-Centric vs Model-Centric in Computer Vision i Göteborg

Thesis Worker at Volvo Cars

Welcome to explore the world of Volvo Cars by writing your thesis with us! As a thesis worker in our organization you are supported by a supervisor who follows you during your project. All thesis projects are arranged in business critical areas and therefore you will be able to contribute to our company purpose – providing freedom to move in a safe, sustainable and personal way – from day one! 

Evaluate Data-Centric vs Model-Centric Approach in Computer Vision 

Background  

There is now a new trend in AI development, moving from Model-Centric to Data-Centric approach, promoted by Andrew Ng. Data-Centric approach focuses on quality of data instead of fixing the code for the machine learning model.  

At the R&D Analytics & Artificial Intelligence Department, we are currently working on Automatic Sitting Posture Analysis, which is used to detect the position of the seat occupant to analyze whether they are sitting in a good position (or a potentially dangerous one). Many studies have shown that Deep Learning methods have a high accuracy in identifying human postures. We have developed a pipeline of Deep Convolutional Neural Networks running on 3D video data to do this analysis automatically and would like to see how the networks can be made even better. 

Scope  

Given the current data and the current models architectures, we would like to see which of the above approaches can give the most improvement by individually testing both approaches by leaving either the training procedure or the data fixed and trying to improve only the other. Although the current models are already relatively good, we expect that there is still plenty of room for performance improvements on both approaches, e.g. 

Data-Centric (fixed model, change data) 

• Better train-validation split 
• Collect more examples (There is still more collected data that is not yet annotated) 
• Fix ambiguous labels (the annotation instructions can be improved, some keypoints are inconsistently annotated) 
• Data augmentation (Data augmentation methods could be explored better) 

Model-Centric (fixed data, change model) 

• Current models are only using RGB information without considering the depth information recorded 
• Hyper parameters can be improved 
• Other training methods and cost functions can be explored 
• Other model architectures can be explored 

Together the thesis workers should work to improve the model performance by the ideas above as well as their own ideas, then categorize by, and benchmark the two approaches.  

Do you fit the profile  

We see that you’re studying statistics, computer science, applied mathematics with a focus on machine learning. You have advanced programming skills, particularly in python, strong analytical ability to solve complex problems and you are great with communication and collaboration.  You are available to work on site at least one day a week in Torslanda, Gothenburg. 

Duration  

• Start in January 2022 
• 30 points ECTS (academic credits) if in agreement with your Thesis Advisor in University  
• 2 students  

How to learn more and apply  
Selection will be ongoing during the application period, so do not hesitate to send in your application. Attach your Cv and personal letter stating your interests within the given area and your thoughts and credentials. Apply as soon as possible but no later than October 31st. Please note that applications via email will not be accepted.  

For questions about the thesis work please contact Sheng Huang at sheng.huang@volvocars.com.

Who are we?

Everything we do starts with people. Our purpose is to provide freedom to move, in a personal, sustainable and safe way. We are committed to simplifying our customers’ lives by offering better technology solutions that improve their impact on the world and bringing the most advanced mobility innovations to protect them, their loved ones and the people around them. 

Volvo Cars’ continued success is the result of a collaborative, diverse, and inclusive working environment. The people of Volvo Cars are committed to making a difference in our world. Today, we are one of the most well-known and respected car brands, with over 40,000 employees across the globe. We believe in bringing out the best in each other and harnessing the true power of people. At Volvo Cars your career is designed around your talents and aspirations so you can reach your full potential. Join us on a journey of a lifetime as we create safety, autonomous driving and electrification technologies of tomorrow.