Radu Alexandru Rosu

I'm a PhD student at the University of Bonn (Germany) where I work on problems at the intersection of traditional 3D processing and deep learning. My interests are in 3D reconstruction, novel-view rendering, implicit representations and semantic segmentation.

During my PhD, I interned at Facebook Reality Labs: at Pittsburgh (supervised by Giljoo Nam). During the internship I worked on 3D hair reconstruction from images which culminated with the article Neural Strands presented at ECCV 2022 (Tel Aviv).

I received my Masters from the University of Bonn in September 2018, where I was advised by Sven Behnke. My master thesis was on Semi-Supervised Semantic Mapping through Label Propagation with Semantic Texture Meshes

I obtained my bachelor degree from University of Salamanca (Spain) in 2015, where I majored in Computer Science. I worked with Iván Álvarez Navia on "Reconstruction of 3D Figures from Computerized Tomography".

I grew up in Targoviste (Romania) and later in Aranda de Duero (Spain).

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

* denotes equal contribution co-authorship

We introuce a novel dataset for 3D hairstyles containing 40K samples. Using the dataset we train a diffusion model to recover 3D hair strands from a single RGB image in as little as 3s.

We introduce the permutohedral lattice in the context of neural surface reconstruction and recover geometry and color of a scene given posed RGB images in as little as 30min.

We propose a end-to-end learning system to recover strand-accurate hair geometry and appearance from multi-view images.

We propose a real-time novel-view synthesis method that runs at interactive speeds and generalizes to novel objects after training on a general dataset.

Real-time physically-based renderer with an emphasis on ease-of-use. Actively used for synthetic dataset creation, data visualization, figure creation, etc. Supports deffered rendering, ambient occlusion, bloom, image-based lighting, shader hotloading and various other features.

We create a system for real-time semantic segmentation of general 3D point clouds by embedding points into a permutohedral lattice where convolutions are defined using custom and highly optimized CUDA kernels.

Based on laser data, RGB and thermal images we reconstruct 3D scenes as a textured mesh colored by RGB and thermal information on which we detect potential fires for the purpose of firefigher intervention.

We reconstruct textured meshes from depth data and fuse semantic information into sparse multi-channel textures.

You've probably seen this website template before, thanks to Jon Barron.
Last updated May 2020.