Mattia D'Urso

PhD candidate 3D Computer Vision

I am a PhD student in Prof. Friedrich Fraundorfer's group at the Institute for Visual Computing (IVC) at Graz University of Technology, Austria.

I am currently visiting the Computer Vision and Geometry Group at ETH Zürich, supervised by Prof. Marc Pollefeys and Dr. Daniel Barath.

Previously, I gained professional experience at the Canva office in Vienna and at Infineon Technologies in Villach, Austria.

I hold a joint Master's double degree in Computer Science from the University of Klagenfurt and the University of Udine where I also completed my Bachelor's degree in the same field.

Research

My research focuses on 3D Computer Vision and Machine Learning, with a specific interest in Pose Estimation and Structure from Motion (SfM).

EPO: Boosting 3D Foundation Models with Edge-based Pose Optimization

Mattia D'Urso¹ Christian Sormann² Mattia Rossi² Friedrich Fraundorfer¹

¹Graz University of Technology ²Sony Europe

ECCV 2026

EPO is an optimization framework specifically designed to enhance Structure-from-Motion reconstructions generated by 3D foundation models, without the need for explicit feature tracks. We use Gradient Descent to learn a per-scene pose-MLP and a set of weights improving camera parameters and geometry. We demonstrate that EPO achieves geometric precision comparable to, or even exceeding, traditional Bundle Adjustment, while reducing runtime by up to 80% and operating efficiently on consumer-grade hardware. Webpage is available here.

TerraSky3D: Multi-View Reconstructions of European Landmarks in 4K

Mattia D'Urso¹ Yuxi Hu¹ Christian Sormann² Mattia Rossi² Friedrich Fraundorfer¹

¹Graz University of Technology ²Sony Europe

CVPRW 2026 🤗 Dataset 🥈 Best Paper Award (Runner-Up)

TerraSky3D is a high-resolution 4K dataset of European landmarks comprising 50,000 images across 150 scenes, specifically designed to bridge the modality gap between ground-level and aerial perspectives for 3D reconstruction. It provides curated camera poses and dense depth maps. Project page is available here.

SANDesc: A Streamlined Attention-Based Network for Descriptor Extraction

Mattia D'Urso¹ Emanuele Santellani¹ Christian Sormann² Mattia Rossi² Andreas Kuhn² Friedrich Fraundorfer¹

¹Graz University of Technology ²Sony Europe

3DV 2026

SANDesc is a descriptor module that can be trained on top of any existing keypoint detector. It significantly improves pose estimation performance on high-resolution images. Project page is available here.

A Lod of Gaussians: Unified Training and Rendering for Ultra-Large Scale Reconstruction with External Memory

Felix Windisch¹ Thomas Köhler¹ Lukas Radl¹ Mattia D'Urso¹ Michael Steiner¹ Dieter Schmalstieg^1,2 Markus Steinberger^1,3

¹Graz University of Technology ²University of Stuttgart ³Huawei Technologies

SIGGRAPH 2026

A LoD of Gaussians enables the reconstruction and interactive visualization of ultra-large-scale scenes by replacing traditional scene partitioning with a dynamic, out-of-core Level-of-Detail framework. Project page is available here.

PyTorchGeoNodes: Enabling Differentiable Shape Programs for 3D Shape Reconstruction

Siniša Šteković¹ Arslan Artykov¹ Stefan Ainetter² Mattia D'Urso² Friedrich Fraundorfer²

¹École des Ponts ParisTech ²Graz University of Technology

CVPR 2025

PyTorchGeoNodes is a differentiable module that enables reconstruction of 3D objects and their semantic parameters using interpretable shape programs. Project page is available here.