Andrea Ramazzina

Since 2021, I'm a PhD student at Mercedes-Benz and Technical University of Munich, supervised by Federico Tombari, Felix Heide and Mario Bijelic. My main research focus is on understanding the world through generative models.

Before that, from 2019 to 2021 I worked at Volvo Autonomous Solutions developing their V&V analytics and other data science projects. Here, I was also part of the Engineering Graduate Program 2019.

I have a MS in Physics of Complex Systems from Chalmers University of Technology where I worked with Fredrik Kahl, and a double BS degree from Politecnico di Milano and Tongji University. During my studies I was recipient of the PoliTong Scholarship (2015), Adlerbert Hospitality Foundation Scholarship (2018) and Taiwan Elite Internship Program Scholarship (2018).

Email  /  CV  /  Scholar  /  Linkedin  /  Github

Research

I'm interested in computer vision, machine learning, optimization, and complex adaptive systems.

* denotes equal contribution.

We present a method to directly generate large-scale 3D driving scenes with accurate geometry, allowing for causal novel view synthesis with object permanence and explicit 3D geometry estimation.

We propose a training-free method to accelerate multimodal reasoning in vision-language models by removing vision tokens after a few decoding steps—significantly reducing memory and computation while maintaining reasoning accuracy.

We introduce a neural scene reconstruction method operating on raw transient measures of outdoor in-the-wild captures to accurately reconstruct the underlying scene geometry and properties.

We introduce a novel evaluation benchmark that extends depth evaluation up to 200 meters under clear, foggy, and rainy conditions, and evaluate state-of-the-art perception models on it.

We propose a neural rendering approach that seamlessly incorporates time-gated capture and illumination. Our method exploits the intrinsic depth cues in the gated videos, achieving precise and dense geometry reconstruction irrespective of ambient illumination conditions.

We introduce HINT, a novel method able to learn a robust representation of a human captured only in a limited range of views augmenting existing sequences with novel views of the person and poses.

We introduce Radar Fields - a neural scene reconstruction method designed for active radar imagers. Our approach unites an explicit, physics-informed sensor model with an implicit neural geometry and reflectance model to directly synthesize raw radar measurements and extract scene occupancy.

We learn a foggy scene from a single video, disjointly representing scattering volume and scene objects through physics-inspired losses. This allows for accurate rendering of novel views with full control over the scattering volume, as well as a more accurate estimation of the 3D scene.

We propose Gated Stereo, a high-resolution and long-range depth estimation technique that operates on active gated stereo images.

We present an improved solution to calculate the so called air-light integral that can be evaluated fast and robustly for an isotropic point source in homogeneous media.

Previously Supervised Students

• Alessandro Sanvito - M.Sc - Now at Optiver
• Tina Höflich - B.Sc - Now at Mercedes-Benz

Design and source code from Jon Barron's website