A probabilistic method combining generative models with measurement consistency for 3D shape reconstruction, traversing Langevin dynamics from a diffusion model while preserving measurement consistency at every step.
Linus Härenstam-Nielsen
I am a PhD student at the Computer Vision Group supervised by Prof. Daniel Cremers at the Technical University of Munich.
Before the PhD, I obtained a M.Sc. degree in Systems control and Robotics from KTH, Royal Institute of Technology (Stockholm, Sweden). I also spent roughly 3 years as a computer vision engineer at Univrses working on visual odometry and deep learning for autonomous driving.
Research
My research focuses on 3D reconstruction from noisy and incomplete sensor data. Using a mix of optimization-based and learning-based methods. Below is a list of selected publications, for the full list see my Google Scholar.
We decouple distortion calibration from 3D reconstruction by working in projective space, averaging multiple pairwise distortion estimates into a single consistent camera model without requiring 3D points or full bundle adjustment.
We improve the robustness of surface reconstruction methods by optimizing a symmetric differentiable Chamfer distance for implicit surfaces.
We develop a certifiably optimal method for triangulating noisy observations with outliers.
We develop a globally optimal solution to the noisy hand-eye calibration problem, using a dual quaternion formulation.
Software
Python library for making 3D plots with Blender. Like Matplotlib for 3D visualization, but using Blender's rendering engine. Install with pip install blender-plots.