The S23DR competition focused on reconstructing a structured 3D model of a house roof from a series of photographs taken from the ground. The goal was not just to create a visually impressive model, but an accurate and measurable description of the roof geometry in the form of peaks and edges, which can be further used in practice.
“Practically, it can be imagined that a person walks around the house and takes a picture of it with a mobile phone from several angles, and the system creates an accurate 3D model of the roof from the photos. Such a model can then be used, for example, for planning the installation of solar panels, calculating the roof area, reconstruction or insurance purposes,” explains Ing. Jan Škvrna.
Artificial intelligence fills in the missing information
The task was not easy. The input data contained only a limited number of spatial points obtained from the photographs. In addition, part of the roof is not visible from the ground and the data contains noise and inaccuracies.
“First, the spatial structure of the scene is estimated from the photographs, and a sparse and inaccurate 3D model is created. However, this is not enough for practical use. This is where our artificial intelligence model comes in, which learns to estimate the actual shape of the roof from incomplete data and creates its structured model,” describes Ing. Miroslav Purkrábek.
The competition was not an academic exercise without practical impact. It was sponsored by the company Hover, which uses similar technologies with its customers and has long been looking for ways to further increase their accuracy. Therefore, the participants solved a real problem from practice – how to reconstruct the actual shape of the roof as accurately as possible from a limited number of photographs. It is precisely such competitions that allow companies to gain inspiration from the latest academic research and for researchers to verify their methods on problems that have direct commercial use.
Drama until the last seconds
The competition was attended by approximately thirty competitors from all over the world. The Czech team’s biggest competitor was Lund University from Sweden.
“The last few days were very exciting. We knew that Lund University had an exceptionally strong solution. When we started sending our best models a few days before the deadline, we found out in the public ranking that we were very close to the top. However, our competitors were also improving their results,” recalls Jan Škvrna.
In the end, literally thousandths of a point were decisive. Even more important, however, was the result on the hidden test dataset, which the organizers used for the final evaluation. It was here that the main advantage of the solution developed at the FEE CTU became apparent – its ability to function correctly even on previously unseen data.
“The results were almost indistinguishable on the public data. The decisive factor was the hidden data set, which verified whether the model could correctly reconstruct new houses and situations that it had not encountered during training,” adds Miroslav Purkrábek.
The organizers evaluated not only the accuracy of the reconstruction, but also the computational complexity. The entire test dataset had to be processed within the prescribed time limit of two hours. “It wasn’t just about creating the largest possible model. The combination of accuracy, speed and robustness was important. Our solution can reconstruct a house in a few seconds and it also worked on a regular laptop,” says Jan Škvrna.
Interestingly, the winning team from FEE CTU and the second team from Lund University achieved almost identical results using completely different technical approaches.
VRG’s third victory in the third year of the competition
The success follows previous victories by members of the Visual Recognition Group in this competition, which is being held for the third year. Jan Škvrna won last year and Denis Rozumnyi, also associated with VRG, succeeded in the first year. The result thus confirms the group’s long-standing expertise in the field of 3D reconstruction and computer vision.
“The workshop brings together top people from institutions and companies such as TU Munich, ETH Zurich and Google, among others. It is gratifying that our doctoral students have managed to succeed in this competition repeatedly. This shows that this is not a one-time success, but the long-term quality of the research we do in the group,” concludes Prof. Jiří Matas, head of the Visual Recognition Group at the Department of Cybernetics, FEE CTU.
The Urban Scene Modeling (USM3D) workshop, within which the competition took place, brings together leading world experts in 3D reconstruction, digital city twins, photogrammetry and computer vision. It is held as part of the CVPR 2026 conference, which is one of the most important global events in the field of artificial intelligence and computer vision.