Mehmet Aygün

My research spans machine learning and computer vision, with a focus on generative and self-supervised methods that reduce supervision and improve robustness in real-world settings. I received my Ph.D. from the University of Edinburgh under Oisin Mac Aodha, and previously worked with Laura Leal-Taixé at the Technical University of Munich.

Currently, I am a Research Scientist at GenBio AI, building large multimodal models for biology at the cell and tissue level. I develop foundational representation models and generative models to enable robust biological insight and scalable discovery.

CV  /  Google Scholar  /  Bluesky

• Sep 2025, I joined GenBio AI as a Research Scientist, in beautiful Paris!!
• Feb 2025, DepthCues got accepted to CVPR 2025.
• Feb 2024, SAOR got accepted to CVPR 2024.
• Aug 2023, I joined Meta AI as a Research Scientist Intern.
• Jul 2022, Our paper about unsupervised semantic correspondence is accepted to ECCV, 2022.
• Feb 2021, Our paper titled 4D Panoptic Lidar Segmentation is accepted to CVPR, 2021.
• Oct 2017, at ICCV 2017, our paper received Honorable Mention Award from TASK-CV workshop.

I am broadly interested in machine learning and computer vision. My work focuses on developing models that can learn fundamental priors and representations with minimal supervision, particularly through generative and self-supervised approaches. I am also interested in improving robustness and generalization in these systems.

This thesis advances computer vision by improving semantic correspondence, unsupervised 3D shape recovery, and self-supervised learning with 3D priors. Together, these methods enhance model robustness and move machine perception closer to human-level understanding..

A new benchmark designed to evaluate depth cue understanding, with findings across 20 diverse and representative pre-trained vision models.

A new self-supervised approach for estimating shape of highly articulated objects such as animals from single-view images.

A new approach using 3D reconstruction to enhance the robustness of self-supervised vision models.

A novel unsupervised approach for semantic correspondence problem, an evaluation framework along with a metric, and through evaluation of current methods.

A new method and evaluation metric for the task of 4D panoptic segmentation.

An emprical study showing that Heat Kernels can replace Geodesic Matrices for unsupervised dense shape correnspondence problem