On Wednesday, 20 May 2026, the Bruno Kessler Foundation (FBK) in Trento hosted the third edition of the Computer Vision Trento Symposium (CVTS). It was an intense day that brought together over 130 participants, including researchers, students, and industry professionals.
The symposium, organised by Fabio Poiesi, Sergio Povoli, Luigi Riz, and Davide Talon, and sponsored by the IT4LIA AI Factory project, has established itself as the main reference point for the Italian scientific community ahead of the prestigious CVPR 2026 world conference in Denver.
A Path of Exponential Growth
The opening remarks were delivered by Fabio Poiesi, Head of the Technologies of Vision (TeV) Unit at FBK‘s Digital Industry centre. Welcoming the large audience in the Sala Stringa, Poiesi expressed great satisfaction at the participation of representatives from the leading research institutions in north-eastern Italy. The data presented shows a significant growth trend: from 70 registered attendees at the inaugural edition in 2024, to 100 in 2025, reaching a record 140 registrations this year. “The event,” Poiesi stated, “is shaping up as an opportunity to foster collaborations and build networks, allowing authors to present work accepted at CVPR 2026 in Denver ahead of the conference.”
From Open Source Project to World Standard: The Keynote
One of the highlights of the day was the talk by Francesco Siddi, CEO of Blender, entitled “The Making of Blender“, Siddi traced the extraordinary evolution of a project born in 2002 through the efforts of a handful of passionate volunteers, which has today become one of the global standards for 3D modelling and rendering. Siddi highlighted how the open-source approach and community-driven development were the engines behind remarkable successes, citing as a landmark example Flow, the animated film made entirely with Blender and winner of the 2025 Oscar for Best Animated Feature.
Cutting-Edge Research and Industrial Applications
The technical core of the symposium featured the presentation of 17 scientific papers accepted at CVPR 2026, spread across five thematic sessions covering topics from Generative AI to Transfer Learning, from Grounded Reasoning to 3D Representations, and Vision Language Model applications.
Among these contributions, three papers stood out for their strong thematic alignment with the areas of interest of IT4LIA AI Factory’s verticals, while remaining expressions of the frontier research typical of the international academic community. These are advanced solutions exploring still-open problems, often at a different level of technological maturity compared to industrial applications, but anticipating their directions and potential developments.
The first, entitled “Specificity-aware reinforcement learning for fine-grained open-world classification” by Angheben et al., addresses an unresolved limitation in Vision Language Models: the ability to go beyond general classification to capture the semantic specificity of objects. While current models can correctly recognise broad categories such as “bird,” they far more often fail to determine the family or sub-class with precision. The work proposes a reinforcement learning approach to improve this granularity – a line of research that is thematically close to FBK-TeV’s activities in the Agrifood & Agritech vertical, where the goal is to adapt generalist models to industrial and productive contexts, making them reliable tools for concrete applications in the agricultural domain.
A second contribution, “TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation” by Shu et al., explores spatial reasoning applied to Earth observation. The method enables Vision Language Models to interpret satellite images at pixel level and answer natural language queries, such as estimating the extent of built-up areas. Here too, the work sits on a frontier research trajectory that dialogues thematically with IT4LIA AI Factory’s Earth vertical, which is instead focused on enabling operational services and tools for territorial monitoring and applied use of geospatial data.
Finally, “MIRAGE: Model-agnostic Industrial Realistic Anomaly Generation and Evaluation for Visual Anomaly Detection” by Hu et al. proposes a fully automated pipeline for the generation and evaluation of realistic industrial anomalies. The contribution addresses a crucial challenge in Visual Anomaly Detection, namely the scarcity of real defect data, by proposing a generative solution to bridge this gap. Again, the connection is thematic rather than applicative: the research fits within the broader field of AI-based quality control, which in IT4LIA AI Factory’s Manufacturing and Agrifood & Agritech verticals takes on a form oriented towards integration into industrial processes.
Networking and Collaboration
Beyond its scientific and academic value, the CVTS once again proved to be a key moment for networking and strengthening the scientific community. The social moments during coffee breaks and the lunch break — enriched by a poster session to facilitate interaction and the exchange of ideas among attendees — were made possible thanks to the sponsorship of the IT4LIA AI Factory project, and offered an informal yet stimulating space for dialogue. These encounters helped forge new connections between researchers and private sector representatives, creating a direct bridge between cutting-edge theoretical research and the operational needs of the industrial world, in line with the integration objectives promoted by IT4LIA AI Factory’s verticals.
An Event That Renews Itself
At the close of the day, Fabio Poiesi thanked participants and organisers for the strong turnout and the quality of the contributions presented, once again highlighting the symposium’s value as a privileged space for exchange between research and innovation. The event was duly renewed to its next edition, with the hope of continuing to strengthen an increasingly cohesive community capable of translating frontier research results into concrete impacts for the industrial and applied world.
