Do you want to create technology for the next generation of industry and society? Univrses is now looking for a Master Thesis student to join the team in Stockholm!

Univrses is a 3D Computer Vision and Machine Learning company based in Stockholm, creating high-end technologies for autonomous systems. We work in several different areas, but our main focus is on self-driving vehicles, mobile robotics and smart city development. Our team consists of hard-working and friendly people from all over the world with diverse backgrounds and expert knowledge in computer vision, robotics, machine learning, physics, math, software development and more.

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.

Background

Finding correspondences between two images is a fundamental building block of many computer vision applications like camera tracking and 3D mapping. The most common approach to image matching relies on sparse interest points that are matched using high-dimensional representations encoding their local visual appearance. Reliably describing each point is challenging in conditions that exhibit symmetries, weak texture, or appearance changes due to varying viewpoints and lighting. To reject outliers that arise from occlusion and missing points, such representations should also be discriminative. This yields two conflicting objectives, robustness, and uniqueness, that are hard to satisfy.

To address these limitations, SuperGlue introduced a new paradigm – a deep network that considers both images at the same time to jointly match sparse points and reject outliers. It leverages the powerful Transformer model to learn to match challenging image pairs from large datasets.

Univrses need to uniquely identify features across several images captured from different viewpoints and with different lighting/seasons.

Goals

The primary objectives of this research are as follows:

1. Deep feature matching for bounding box matching: How can these point-to-point matches be extended to bounding box to bounding box matching?
2. Pretrained models vs fine-tuned model: What are the performances of the pretrained model on our data? Do we benefit from fine-tuning them on our data?
3. Robustness evaluation: How robust is it to large point of view changes, temporal changes (time of the day, time of the year), repetitive patterns?

Relevant material

• Feature matching:
  • LightGlue: Local Feature Matching at Light Speed
• Feature extractors:
  • SuperPoint: Self-Supervised Interest Point Detection and Description
  • DISK: Learning local features with policy gradient

Easily test different image-matching solutions here: Image Matching Webui - a Hugging Face Space by Realcat

Required background

Applicants are expected to possess a comprehensive understanding of deep learning and computer vision, which will be essential for engaging with the research effectively.

Proficiency in Python coding is a prerequisite for implementing and experimenting with the proposed techniques.

Required application material

• CV
• Cover letter
• Transcripts (with grades)

All application material must be in English.
The work is supposed to start in January 2024

Next step

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