We focus on foundational academic AI research. We publish in open venues, share our code, and push the field forward.
How can AI agents adapt to ever-changing environments? By seamlessly integrating advanced perception, Q&A, and control strategies — alongside techniques like test-time training, model merging, and on-the-fly knowledge updates — we aim to develop robust, lifelong-learning systems that continuously refine their capabilities in response to new rules, tasks, and scenarios.
Text-to-image diffusion models excel at generating high-quality visuals, but reconstructing entire scenarios is challenging due to greater spatial complexity. In this project, we propose a pipeline that synthesizes realistic 3D environments from minimal input, combining texturing and semantic preservation with pathfinding and 3D Gaussian Splatting.
ICCV • 2023
Marc Botet Colomer, Pier Luigi Dovesi, Theodoros Panagiotakopoulos, Joao Frederico Carvalho, Linus Härenstam-Nielsen, Hossein Azizpour, Hedvig Kjellström, Daniel Cremers, Matteo Poggi
The paper explores real-time adaptation strategies for semantic segmentation models, discussing the trade-offs between adaptation and performance in dynamic environments.
ECCV • 2022
Theodoros Panagiotakopoulos, Pier Luigi Dovesi, Linus Härenstam-Nielsen, Matteo Poggi
This work introduces a framework for online domain adaptation in semantic segmentation, allowing models to adapt to continuously changing environments without offline retraining.
ICRA • 2020
Pier Luigi Dovesi, Matteo Poggi, Lorenzo Andraghetti, Miquel Martí, Hedvig Kjellström, Alessandro Pieropan, Stefano Mattoccia
This paper presents a compact and lightweight architecture for real-time semantic stereo matching, enabling efficient inference on embedded devices with minimal accuracy loss.
3DV • 2019
Lorenzo Andraghetti, Panteleimon Myriokefalitakis, Pier Luigi Dovesi, Belén Luque, Matteo Poggi, Alessandro Pieropan, Stefano Mattoccia
The authors propose a method to improve self-supervised monocular depth estimation by integrating traditional visual odometry techniques, achieving better accuracy in depth prediction.