Do you want to create technology for the next generation of industry and society while working on your Master Thesis? Then you’ve come to the right place!

Who are we?

Univrses is a 3D Computer Vision and Machine Learning company, creating high-end technologies for autonomous systems. Our main focus is on smart cities and self-driving vehicles. The Company is based in Stockholm, Sweden, but we work with clients all over the world. Our team now consists of 50 hard-working and friendly people, all working in our bright and spacious facilities by Medborgarplatsen in the heart of Södermalm, Stockholm. We are constantly expanding due to the increasing demand for high-quality technologies in 3D perception and mapping. And that’s where you come in!

Master Thesis at Univrses

Right now, we offer Master students the opportunity to be an integral part of the team while working on their Thesis. You will be working in a fun, stimulating and highly professional work environment together with our awesome team. This is truly a unique chance to work with some of the best scientists and engineers in Computer Vision and Robotics in the world.

The projects: Multi-modal learning meets Domain Adaptation

Cross-city domain shift is the cause of a 25-30% accuracy drop in normal Deep Learning predictors. Changes in weather and lighting conditions might even result in more disruptive effects. The cause of this disruption lies in a gap between the training and deployment domain, we refer to this concept as “domain shift”. Therefore, to provide a robust and accurate output, the predictor would need to be trained on every possible deployment condition, e.g. multiple cities, districts, seasons, kinds of weather, cameras pose and camera intrinsics. This vastly reduces AI solutions' scalability in terms of costs and deployment time. Self-supervised learning Unsupervised Domain Adaptation (UDA) methods proved to be very effective in mitigating (or even solving) these shortcomings. Univrses has been active for several years in the research of UDA methods in Vision Deep Learning, in the form of MSc theses, internal investigations and published conference papers.

Recently, Self-Supervised Learning (SSL) achieved astonishing results, managing to learn from unstructured and unlabelled data with performances competitive to fully supervised methods. Such sharp improvements are shaking the foundations of Deep Learning as we know it.

Several deep learning long-sought goals now seem within’s arm's reach, e.g. open set classification and adaptation, continuously adaptive systems, and multimodal learning with scarce annotations. With the advent of CLIP and other SSL and Online Adaption techniques, we’re seeing the dawn of a new Deep Learning era, where we’re moving away from fixed semantics, fixed datasets and fixed objectives.

Goals

The goal of this project is to leverage the recent advancement in SSL and multimodal learning to find new ways to tackle domain adaptation. This means for you to:

• Perform an extensive literature review on Vision UDA and SSL methods.
• Identify creative and challenging research gaps.
• Design, develop and benchmark your methods.
• Discuss with tech teams how to integrate your results into real products.
• Bonus: Submit your results to top-tier Deep Learning conferences.

Relevant material

Learning Transferable Visual Models From Natural Language Supervision Learning Transferable Visual Models From Natural Language Supervision

Online Domain Adaptation for Semantic Segmentation in Ever-Changing Conditions Online Domain Adaptation for Semantic Segmentation in...

DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation DAFormer: Improving Network Architectures and Training Strategies... 

Next step

Interested in joining Univrses as our new Master Thesis student? Submit your application - All application material must be in English!

Required application material

• CV
• Cover letter
• Transcripts (with grades)