The S23DR competition focused on reconstructing a structured 3D model of a house roof from a series of photographs taken from ground level. The goal was not merely to create a visually appealing model, but to generate an accurate and measurable description of the roof's geometry in the form of vertices and edges that can be directly applied in real-world scenarios.
"In practical terms, you can imagine someone walking around a house and taking photos with a smartphone from several angles. The system then creates an accurate 3D model of the roof from those images. Such a model can be used, for example, to plan solar panel installations, calculate roof area, support renovations, or serve insurance purposes," explains Jan Škvrna.
Artificial Intelligence Fills in Missing Information
The task was far from straightforward. The input data consisted only of a limited number of spatial points extracted from photographs. Parts of the roof are not visible from the ground, and the data itself contains noise and inaccuracies.
"First, the spatial structure of the scene is estimated from the photographs, producing a sparse and imprecise 3D model. However, that is not sufficient for practical applications. This is where our artificial intelligence model comes in—it learns to infer the true shape of the roof from incomplete data and generates its structured representation," says Miroslav Purkrábek.
The competition was not an academic exercise without practical impact. It was sponsored by Hover, a company that deploys similar technologies for its customers and continuously seeks ways to improve their accuracy. Participants therefore tackled a genuine industry challenge: how to reconstruct the true shape of a roof as accurately as possible from a limited number of photographs. Competitions like this allow companies to draw inspiration from cutting-edge academic research while giving researchers the opportunity to validate their methods on problems with direct commercial applications.
A Nail-Biting Finish
Approximately thirty teams and individuals from around the world participated in the competition. The Czech team's strongest competitor was Lund University in Sweden.
"The final days were extremely tense. We knew that Lund University had developed an exceptionally strong solution. When we began submitting our best models a few days before the deadline, the public leaderboard showed that we were very close to the top. At the same time, our competitors kept improving their results," recalls Jan Škvrna.
In the end, the outcome was decided by mere thousandths of a point. Even more important, however, was performance on a hidden test dataset used by the organizers for the final evaluation. This is where the key advantage of the CTU FEE solution became evident—its ability to generalize effectively to previously unseen data.
"On the public dataset, the results were almost indistinguishable. The deciding factor was the hidden dataset, which tested whether the model could correctly reconstruct entirely new houses and situations it had never encountered during training," adds Miroslav Purkrábek.
The organizers evaluated not only reconstruction accuracy but also computational efficiency. The entire test dataset had to be processed within a strict two-hour time limit.
"It wasn't just about building the largest model possible. What mattered was the combination of accuracy, speed, and robustness. Our solution can reconstruct a single house in just a few seconds and was able to run on a standard laptop," says Jan Škvrna.
Interestingly, the winning CTU FEE team and the runner-up team from Lund University achieved nearly identical results using entirely different technical approaches.
VRG's Third Victory in the Competition's Third Edition
This success follows previous victories by members of the Visual Recognition Group in the same competition, which has now been held for the third consecutive year. Jan Škvrna also won last year, while the inaugural edition was won by Denis Rozumnyi, another researcher affiliated with VRG. The result further confirms the group's long-standing expertise in 3D reconstruction and computer vision.
"The workshop brings together leading researchers from institutions and companies such as TU Munich, ETH Zurich, and Google, among others. It is gratifying to see our PhD students succeeding in this competition repeatedly. It demonstrates that this is not a one-off achievement but rather the result of consistently high-quality research carried out within the group," concludes Professor Jiří Matas, Head of the Visual Recognition Group at the Department of Cybernetics, CTU FEE.
The Workshop Urban Scene Modeling (USM3D), within which the competition took place, brings together leading international experts in 3D reconstruction, digital twins of cities, photogrammetry, and computer vision. It is held as part of CVPR 2026, one of the world's most important events in artificial intelligence and computer vision.
Photo credits: Petr Neugebauer
